Actual source code: mpiaij.c
petsc-3.11.4 2019-09-28
3: #include <../src/mat/impls/aij/mpi/mpiaij.h>
4: #include <petsc/private/vecimpl.h>
5: #include <petsc/private/vecscatterimpl.h>
6: #include <petsc/private/isimpl.h>
7: #include <petscblaslapack.h>
8: #include <petscsf.h>
10: /*MC
11: MATAIJ - MATAIJ = "aij" - A matrix type to be used for sparse matrices.
13: This matrix type is identical to MATSEQAIJ when constructed with a single process communicator,
14: and MATMPIAIJ otherwise. As a result, for single process communicators,
15: MatSeqAIJSetPreallocation is supported, and similarly MatMPIAIJSetPreallocation() is supported
16: for communicators controlling multiple processes. It is recommended that you call both of
17: the above preallocation routines for simplicity.
19: Options Database Keys:
20: . -mat_type aij - sets the matrix type to "aij" during a call to MatSetFromOptions()
22: Developer Notes:
23: Subclasses include MATAIJCUSP, MATAIJCUSPARSE, MATAIJPERM, MATAIJSELL, MATAIJMKL, MATAIJCRL, and also automatically switches over to use inodes when
24: enough exist.
26: Level: beginner
28: .seealso: MatCreateAIJ(), MatCreateSeqAIJ(), MATSEQAIJ, MATMPIAIJ
29: M*/
31: /*MC
32: MATAIJCRL - MATAIJCRL = "aijcrl" - A matrix type to be used for sparse matrices.
34: This matrix type is identical to MATSEQAIJCRL when constructed with a single process communicator,
35: and MATMPIAIJCRL otherwise. As a result, for single process communicators,
36: MatSeqAIJSetPreallocation() is supported, and similarly MatMPIAIJSetPreallocation() is supported
37: for communicators controlling multiple processes. It is recommended that you call both of
38: the above preallocation routines for simplicity.
40: Options Database Keys:
41: . -mat_type aijcrl - sets the matrix type to "aijcrl" during a call to MatSetFromOptions()
43: Level: beginner
45: .seealso: MatCreateMPIAIJCRL,MATSEQAIJCRL,MATMPIAIJCRL, MATSEQAIJCRL, MATMPIAIJCRL
46: M*/
48: PetscErrorCode MatSetBlockSizes_MPIAIJ(Mat M, PetscInt rbs, PetscInt cbs)
49: {
51: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
54: if (mat->A) {
55: MatSetBlockSizes(mat->A,rbs,cbs);
56: MatSetBlockSizes(mat->B,rbs,1);
57: }
58: return(0);
59: }
61: PetscErrorCode MatFindNonzeroRows_MPIAIJ(Mat M,IS *keptrows)
62: {
63: PetscErrorCode ierr;
64: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)M->data;
65: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data;
66: Mat_SeqAIJ *b = (Mat_SeqAIJ*)mat->B->data;
67: const PetscInt *ia,*ib;
68: const MatScalar *aa,*bb;
69: PetscInt na,nb,i,j,*rows,cnt=0,n0rows;
70: PetscInt m = M->rmap->n,rstart = M->rmap->rstart;
73: *keptrows = 0;
74: ia = a->i;
75: ib = b->i;
76: for (i=0; i<m; i++) {
77: na = ia[i+1] - ia[i];
78: nb = ib[i+1] - ib[i];
79: if (!na && !nb) {
80: cnt++;
81: goto ok1;
82: }
83: aa = a->a + ia[i];
84: for (j=0; j<na; j++) {
85: if (aa[j] != 0.0) goto ok1;
86: }
87: bb = b->a + ib[i];
88: for (j=0; j <nb; j++) {
89: if (bb[j] != 0.0) goto ok1;
90: }
91: cnt++;
92: ok1:;
93: }
94: MPIU_Allreduce(&cnt,&n0rows,1,MPIU_INT,MPI_SUM,PetscObjectComm((PetscObject)M));
95: if (!n0rows) return(0);
96: PetscMalloc1(M->rmap->n-cnt,&rows);
97: cnt = 0;
98: for (i=0; i<m; i++) {
99: na = ia[i+1] - ia[i];
100: nb = ib[i+1] - ib[i];
101: if (!na && !nb) continue;
102: aa = a->a + ia[i];
103: for (j=0; j<na;j++) {
104: if (aa[j] != 0.0) {
105: rows[cnt++] = rstart + i;
106: goto ok2;
107: }
108: }
109: bb = b->a + ib[i];
110: for (j=0; j<nb; j++) {
111: if (bb[j] != 0.0) {
112: rows[cnt++] = rstart + i;
113: goto ok2;
114: }
115: }
116: ok2:;
117: }
118: ISCreateGeneral(PetscObjectComm((PetscObject)M),cnt,rows,PETSC_OWN_POINTER,keptrows);
119: return(0);
120: }
122: PetscErrorCode MatDiagonalSet_MPIAIJ(Mat Y,Vec D,InsertMode is)
123: {
124: PetscErrorCode ierr;
125: Mat_MPIAIJ *aij = (Mat_MPIAIJ*) Y->data;
126: PetscBool cong;
129: MatHasCongruentLayouts(Y,&cong);
130: if (Y->assembled && cong) {
131: MatDiagonalSet(aij->A,D,is);
132: } else {
133: MatDiagonalSet_Default(Y,D,is);
134: }
135: return(0);
136: }
138: PetscErrorCode MatFindZeroDiagonals_MPIAIJ(Mat M,IS *zrows)
139: {
140: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)M->data;
142: PetscInt i,rstart,nrows,*rows;
145: *zrows = NULL;
146: MatFindZeroDiagonals_SeqAIJ_Private(aij->A,&nrows,&rows);
147: MatGetOwnershipRange(M,&rstart,NULL);
148: for (i=0; i<nrows; i++) rows[i] += rstart;
149: ISCreateGeneral(PetscObjectComm((PetscObject)M),nrows,rows,PETSC_OWN_POINTER,zrows);
150: return(0);
151: }
153: PetscErrorCode MatGetColumnNorms_MPIAIJ(Mat A,NormType type,PetscReal *norms)
154: {
156: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)A->data;
157: PetscInt i,n,*garray = aij->garray;
158: Mat_SeqAIJ *a_aij = (Mat_SeqAIJ*) aij->A->data;
159: Mat_SeqAIJ *b_aij = (Mat_SeqAIJ*) aij->B->data;
160: PetscReal *work;
163: MatGetSize(A,NULL,&n);
164: PetscCalloc1(n,&work);
165: if (type == NORM_2) {
166: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
167: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]*a_aij->a[i]);
168: }
169: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
170: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]*b_aij->a[i]);
171: }
172: } else if (type == NORM_1) {
173: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
174: work[A->cmap->rstart + a_aij->j[i]] += PetscAbsScalar(a_aij->a[i]);
175: }
176: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
177: work[garray[b_aij->j[i]]] += PetscAbsScalar(b_aij->a[i]);
178: }
179: } else if (type == NORM_INFINITY) {
180: for (i=0; i<a_aij->i[aij->A->rmap->n]; i++) {
181: work[A->cmap->rstart + a_aij->j[i]] = PetscMax(PetscAbsScalar(a_aij->a[i]), work[A->cmap->rstart + a_aij->j[i]]);
182: }
183: for (i=0; i<b_aij->i[aij->B->rmap->n]; i++) {
184: work[garray[b_aij->j[i]]] = PetscMax(PetscAbsScalar(b_aij->a[i]),work[garray[b_aij->j[i]]]);
185: }
187: } else SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_ARG_WRONG,"Unknown NormType");
188: if (type == NORM_INFINITY) {
189: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)A));
190: } else {
191: MPIU_Allreduce(work,norms,n,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)A));
192: }
193: PetscFree(work);
194: if (type == NORM_2) {
195: for (i=0; i<n; i++) norms[i] = PetscSqrtReal(norms[i]);
196: }
197: return(0);
198: }
200: PetscErrorCode MatFindOffBlockDiagonalEntries_MPIAIJ(Mat A,IS *is)
201: {
202: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
203: IS sis,gis;
204: PetscErrorCode ierr;
205: const PetscInt *isis,*igis;
206: PetscInt n,*iis,nsis,ngis,rstart,i;
209: MatFindOffBlockDiagonalEntries(a->A,&sis);
210: MatFindNonzeroRows(a->B,&gis);
211: ISGetSize(gis,&ngis);
212: ISGetSize(sis,&nsis);
213: ISGetIndices(sis,&isis);
214: ISGetIndices(gis,&igis);
216: PetscMalloc1(ngis+nsis,&iis);
217: PetscMemcpy(iis,igis,ngis*sizeof(PetscInt));
218: PetscMemcpy(iis+ngis,isis,nsis*sizeof(PetscInt));
219: n = ngis + nsis;
220: PetscSortRemoveDupsInt(&n,iis);
221: MatGetOwnershipRange(A,&rstart,NULL);
222: for (i=0; i<n; i++) iis[i] += rstart;
223: ISCreateGeneral(PetscObjectComm((PetscObject)A),n,iis,PETSC_OWN_POINTER,is);
225: ISRestoreIndices(sis,&isis);
226: ISRestoreIndices(gis,&igis);
227: ISDestroy(&sis);
228: ISDestroy(&gis);
229: return(0);
230: }
232: /*
233: Distributes a SeqAIJ matrix across a set of processes. Code stolen from
234: MatLoad_MPIAIJ(). Horrible lack of reuse. Should be a routine for each matrix type.
236: Only for square matrices
238: Used by a preconditioner, hence PETSC_EXTERN
239: */
240: PETSC_EXTERN PetscErrorCode MatDistribute_MPIAIJ(MPI_Comm comm,Mat gmat,PetscInt m,MatReuse reuse,Mat *inmat)
241: {
242: PetscMPIInt rank,size;
243: PetscInt *rowners,*dlens,*olens,i,rstart,rend,j,jj,nz = 0,*gmataj,cnt,row,*ld,bses[2];
245: Mat mat;
246: Mat_SeqAIJ *gmata;
247: PetscMPIInt tag;
248: MPI_Status status;
249: PetscBool aij;
250: MatScalar *gmataa,*ao,*ad,*gmataarestore=0;
253: MPI_Comm_rank(comm,&rank);
254: MPI_Comm_size(comm,&size);
255: if (!rank) {
256: PetscObjectTypeCompare((PetscObject)gmat,MATSEQAIJ,&aij);
257: if (!aij) SETERRQ1(PetscObjectComm((PetscObject)gmat),PETSC_ERR_SUP,"Currently no support for input matrix of type %s\n",((PetscObject)gmat)->type_name);
258: }
259: if (reuse == MAT_INITIAL_MATRIX) {
260: MatCreate(comm,&mat);
261: MatSetSizes(mat,m,m,PETSC_DETERMINE,PETSC_DETERMINE);
262: MatGetBlockSizes(gmat,&bses[0],&bses[1]);
263: MPI_Bcast(bses,2,MPIU_INT,0,comm);
264: MatSetBlockSizes(mat,bses[0],bses[1]);
265: MatSetType(mat,MATAIJ);
266: PetscMalloc1(size+1,&rowners);
267: PetscMalloc2(m,&dlens,m,&olens);
268: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
270: rowners[0] = 0;
271: for (i=2; i<=size; i++) rowners[i] += rowners[i-1];
272: rstart = rowners[rank];
273: rend = rowners[rank+1];
274: PetscObjectGetNewTag((PetscObject)mat,&tag);
275: if (!rank) {
276: gmata = (Mat_SeqAIJ*) gmat->data;
277: /* send row lengths to all processors */
278: for (i=0; i<m; i++) dlens[i] = gmata->ilen[i];
279: for (i=1; i<size; i++) {
280: MPI_Send(gmata->ilen + rowners[i],rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
281: }
282: /* determine number diagonal and off-diagonal counts */
283: PetscMemzero(olens,m*sizeof(PetscInt));
284: PetscCalloc1(m,&ld);
285: jj = 0;
286: for (i=0; i<m; i++) {
287: for (j=0; j<dlens[i]; j++) {
288: if (gmata->j[jj] < rstart) ld[i]++;
289: if (gmata->j[jj] < rstart || gmata->j[jj] >= rend) olens[i]++;
290: jj++;
291: }
292: }
293: /* send column indices to other processes */
294: for (i=1; i<size; i++) {
295: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
296: MPI_Send(&nz,1,MPIU_INT,i,tag,comm);
297: MPI_Send(gmata->j + gmata->i[rowners[i]],nz,MPIU_INT,i,tag,comm);
298: }
300: /* send numerical values to other processes */
301: for (i=1; i<size; i++) {
302: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
303: MPI_Send(gmata->a + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
304: }
305: gmataa = gmata->a;
306: gmataj = gmata->j;
308: } else {
309: /* receive row lengths */
310: MPI_Recv(dlens,m,MPIU_INT,0,tag,comm,&status);
311: /* receive column indices */
312: MPI_Recv(&nz,1,MPIU_INT,0,tag,comm,&status);
313: PetscMalloc2(nz,&gmataa,nz,&gmataj);
314: MPI_Recv(gmataj,nz,MPIU_INT,0,tag,comm,&status);
315: /* determine number diagonal and off-diagonal counts */
316: PetscMemzero(olens,m*sizeof(PetscInt));
317: PetscCalloc1(m,&ld);
318: jj = 0;
319: for (i=0; i<m; i++) {
320: for (j=0; j<dlens[i]; j++) {
321: if (gmataj[jj] < rstart) ld[i]++;
322: if (gmataj[jj] < rstart || gmataj[jj] >= rend) olens[i]++;
323: jj++;
324: }
325: }
326: /* receive numerical values */
327: PetscMemzero(gmataa,nz*sizeof(PetscScalar));
328: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
329: }
330: /* set preallocation */
331: for (i=0; i<m; i++) {
332: dlens[i] -= olens[i];
333: }
334: MatSeqAIJSetPreallocation(mat,0,dlens);
335: MatMPIAIJSetPreallocation(mat,0,dlens,0,olens);
337: for (i=0; i<m; i++) {
338: dlens[i] += olens[i];
339: }
340: cnt = 0;
341: for (i=0; i<m; i++) {
342: row = rstart + i;
343: MatSetValues(mat,1,&row,dlens[i],gmataj+cnt,gmataa+cnt,INSERT_VALUES);
344: cnt += dlens[i];
345: }
346: if (rank) {
347: PetscFree2(gmataa,gmataj);
348: }
349: PetscFree2(dlens,olens);
350: PetscFree(rowners);
352: ((Mat_MPIAIJ*)(mat->data))->ld = ld;
354: *inmat = mat;
355: } else { /* column indices are already set; only need to move over numerical values from process 0 */
356: Mat_SeqAIJ *Ad = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->A->data;
357: Mat_SeqAIJ *Ao = (Mat_SeqAIJ*)((Mat_MPIAIJ*)((*inmat)->data))->B->data;
358: mat = *inmat;
359: PetscObjectGetNewTag((PetscObject)mat,&tag);
360: if (!rank) {
361: /* send numerical values to other processes */
362: gmata = (Mat_SeqAIJ*) gmat->data;
363: MatGetOwnershipRanges(mat,(const PetscInt**)&rowners);
364: gmataa = gmata->a;
365: for (i=1; i<size; i++) {
366: nz = gmata->i[rowners[i+1]]-gmata->i[rowners[i]];
367: MPI_Send(gmataa + gmata->i[rowners[i]],nz,MPIU_SCALAR,i,tag,comm);
368: }
369: nz = gmata->i[rowners[1]]-gmata->i[rowners[0]];
370: } else {
371: /* receive numerical values from process 0*/
372: nz = Ad->nz + Ao->nz;
373: PetscMalloc1(nz,&gmataa); gmataarestore = gmataa;
374: MPI_Recv(gmataa,nz,MPIU_SCALAR,0,tag,comm,&status);
375: }
376: /* transfer numerical values into the diagonal A and off diagonal B parts of mat */
377: ld = ((Mat_MPIAIJ*)(mat->data))->ld;
378: ad = Ad->a;
379: ao = Ao->a;
380: if (mat->rmap->n) {
381: i = 0;
382: nz = ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
383: nz = Ad->i[i+1] - Ad->i[i]; PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
384: }
385: for (i=1; i<mat->rmap->n; i++) {
386: nz = Ao->i[i] - Ao->i[i-1] - ld[i-1] + ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar)); ao += nz; gmataa += nz;
387: nz = Ad->i[i+1] - Ad->i[i]; PetscMemcpy(ad,gmataa,nz*sizeof(PetscScalar)); ad += nz; gmataa += nz;
388: }
389: i--;
390: if (mat->rmap->n) {
391: nz = Ao->i[i+1] - Ao->i[i] - ld[i]; PetscMemcpy(ao,gmataa,nz*sizeof(PetscScalar));
392: }
393: if (rank) {
394: PetscFree(gmataarestore);
395: }
396: }
397: MatAssemblyBegin(mat,MAT_FINAL_ASSEMBLY);
398: MatAssemblyEnd(mat,MAT_FINAL_ASSEMBLY);
399: return(0);
400: }
402: /*
403: Local utility routine that creates a mapping from the global column
404: number to the local number in the off-diagonal part of the local
405: storage of the matrix. When PETSC_USE_CTABLE is used this is scalable at
406: a slightly higher hash table cost; without it it is not scalable (each processor
407: has an order N integer array but is fast to acess.
408: */
409: PetscErrorCode MatCreateColmap_MPIAIJ_Private(Mat mat)
410: {
411: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
413: PetscInt n = aij->B->cmap->n,i;
416: if (!aij->garray) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_PLIB,"MPIAIJ Matrix was assembled but is missing garray");
417: #if defined(PETSC_USE_CTABLE)
418: PetscTableCreate(n,mat->cmap->N+1,&aij->colmap);
419: for (i=0; i<n; i++) {
420: PetscTableAdd(aij->colmap,aij->garray[i]+1,i+1,INSERT_VALUES);
421: }
422: #else
423: PetscCalloc1(mat->cmap->N+1,&aij->colmap);
424: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N+1)*sizeof(PetscInt));
425: for (i=0; i<n; i++) aij->colmap[aij->garray[i]] = i+1;
426: #endif
427: return(0);
428: }
430: #define MatSetValues_SeqAIJ_A_Private(row,col,value,addv,orow,ocol) \
431: { \
432: if (col <= lastcol1) low1 = 0; \
433: else high1 = nrow1; \
434: lastcol1 = col;\
435: while (high1-low1 > 5) { \
436: t = (low1+high1)/2; \
437: if (rp1[t] > col) high1 = t; \
438: else low1 = t; \
439: } \
440: for (_i=low1; _i<high1; _i++) { \
441: if (rp1[_i] > col) break; \
442: if (rp1[_i] == col) { \
443: if (addv == ADD_VALUES) ap1[_i] += value; \
444: else ap1[_i] = value; \
445: goto a_noinsert; \
446: } \
447: } \
448: if (value == 0.0 && ignorezeroentries && row != col) {low1 = 0; high1 = nrow1;goto a_noinsert;} \
449: if (nonew == 1) {low1 = 0; high1 = nrow1; goto a_noinsert;} \
450: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
451: MatSeqXAIJReallocateAIJ(A,am,1,nrow1,row,col,rmax1,aa,ai,aj,rp1,ap1,aimax,nonew,MatScalar); \
452: N = nrow1++ - 1; a->nz++; high1++; \
453: /* shift up all the later entries in this row */ \
454: for (ii=N; ii>=_i; ii--) { \
455: rp1[ii+1] = rp1[ii]; \
456: ap1[ii+1] = ap1[ii]; \
457: } \
458: rp1[_i] = col; \
459: ap1[_i] = value; \
460: A->nonzerostate++;\
461: a_noinsert: ; \
462: ailen[row] = nrow1; \
463: }
465: #define MatSetValues_SeqAIJ_B_Private(row,col,value,addv,orow,ocol) \
466: { \
467: if (col <= lastcol2) low2 = 0; \
468: else high2 = nrow2; \
469: lastcol2 = col; \
470: while (high2-low2 > 5) { \
471: t = (low2+high2)/2; \
472: if (rp2[t] > col) high2 = t; \
473: else low2 = t; \
474: } \
475: for (_i=low2; _i<high2; _i++) { \
476: if (rp2[_i] > col) break; \
477: if (rp2[_i] == col) { \
478: if (addv == ADD_VALUES) ap2[_i] += value; \
479: else ap2[_i] = value; \
480: goto b_noinsert; \
481: } \
482: } \
483: if (value == 0.0 && ignorezeroentries) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
484: if (nonew == 1) {low2 = 0; high2 = nrow2; goto b_noinsert;} \
485: if (nonew == -1) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", orow, ocol); \
486: MatSeqXAIJReallocateAIJ(B,bm,1,nrow2,row,col,rmax2,ba,bi,bj,rp2,ap2,bimax,nonew,MatScalar); \
487: N = nrow2++ - 1; b->nz++; high2++; \
488: /* shift up all the later entries in this row */ \
489: for (ii=N; ii>=_i; ii--) { \
490: rp2[ii+1] = rp2[ii]; \
491: ap2[ii+1] = ap2[ii]; \
492: } \
493: rp2[_i] = col; \
494: ap2[_i] = value; \
495: B->nonzerostate++; \
496: b_noinsert: ; \
497: bilen[row] = nrow2; \
498: }
500: PetscErrorCode MatSetValuesRow_MPIAIJ(Mat A,PetscInt row,const PetscScalar v[])
501: {
502: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)A->data;
503: Mat_SeqAIJ *a = (Mat_SeqAIJ*)mat->A->data,*b = (Mat_SeqAIJ*)mat->B->data;
505: PetscInt l,*garray = mat->garray,diag;
508: /* code only works for square matrices A */
510: /* find size of row to the left of the diagonal part */
511: MatGetOwnershipRange(A,&diag,0);
512: row = row - diag;
513: for (l=0; l<b->i[row+1]-b->i[row]; l++) {
514: if (garray[b->j[b->i[row]+l]] > diag) break;
515: }
516: PetscMemcpy(b->a+b->i[row],v,l*sizeof(PetscScalar));
518: /* diagonal part */
519: PetscMemcpy(a->a+a->i[row],v+l,(a->i[row+1]-a->i[row])*sizeof(PetscScalar));
521: /* right of diagonal part */
522: PetscMemcpy(b->a+b->i[row]+l,v+l+a->i[row+1]-a->i[row],(b->i[row+1]-b->i[row]-l)*sizeof(PetscScalar));
523: return(0);
524: }
526: PetscErrorCode MatSetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt im[],PetscInt n,const PetscInt in[],const PetscScalar v[],InsertMode addv)
527: {
528: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
529: PetscScalar value;
531: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
532: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
533: PetscBool roworiented = aij->roworiented;
535: /* Some Variables required in the macro */
536: Mat A = aij->A;
537: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
538: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
539: MatScalar *aa = a->a;
540: PetscBool ignorezeroentries = a->ignorezeroentries;
541: Mat B = aij->B;
542: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
543: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
544: MatScalar *ba = b->a;
546: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
547: PetscInt nonew;
548: MatScalar *ap1,*ap2;
551: for (i=0; i<m; i++) {
552: if (im[i] < 0) continue;
553: #if defined(PETSC_USE_DEBUG)
554: if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
555: #endif
556: if (im[i] >= rstart && im[i] < rend) {
557: row = im[i] - rstart;
558: lastcol1 = -1;
559: rp1 = aj + ai[row];
560: ap1 = aa + ai[row];
561: rmax1 = aimax[row];
562: nrow1 = ailen[row];
563: low1 = 0;
564: high1 = nrow1;
565: lastcol2 = -1;
566: rp2 = bj + bi[row];
567: ap2 = ba + bi[row];
568: rmax2 = bimax[row];
569: nrow2 = bilen[row];
570: low2 = 0;
571: high2 = nrow2;
573: for (j=0; j<n; j++) {
574: if (roworiented) value = v[i*n+j];
575: else value = v[i+j*m];
576: if (in[j] >= cstart && in[j] < cend) {
577: col = in[j] - cstart;
578: nonew = a->nonew;
579: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue;
580: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
581: } else if (in[j] < 0) continue;
582: #if defined(PETSC_USE_DEBUG)
583: else if (in[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);
584: #endif
585: else {
586: if (mat->was_assembled) {
587: if (!aij->colmap) {
588: MatCreateColmap_MPIAIJ_Private(mat);
589: }
590: #if defined(PETSC_USE_CTABLE)
591: PetscTableFind(aij->colmap,in[j]+1,&col);
592: col--;
593: #else
594: col = aij->colmap[in[j]] - 1;
595: #endif
596: if (col < 0 && !((Mat_SeqAIJ*)(aij->B->data))->nonew) {
597: MatDisAssemble_MPIAIJ(mat);
598: col = in[j];
599: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
600: B = aij->B;
601: b = (Mat_SeqAIJ*)B->data;
602: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j; ba = b->a;
603: rp2 = bj + bi[row];
604: ap2 = ba + bi[row];
605: rmax2 = bimax[row];
606: nrow2 = bilen[row];
607: low2 = 0;
608: high2 = nrow2;
609: bm = aij->B->rmap->n;
610: ba = b->a;
611: } else if (col < 0) {
612: if (1 == ((Mat_SeqAIJ*)(aij->B->data))->nonew) {
613: PetscInfo3(mat,"Skipping of insertion of new nonzero location in off-diagonal portion of matrix %g(%D,%D)\n",(double)PetscRealPart(value),im[i],in[j]);
614: } else SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Inserting a new nonzero at global row/column (%D, %D) into matrix", im[i], in[j]);
615: }
616: } else col = in[j];
617: nonew = b->nonew;
618: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
619: }
620: }
621: } else {
622: if (mat->nooffprocentries) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Setting off process row %D even though MatSetOption(,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) was set",im[i]);
623: if (!aij->donotstash) {
624: mat->assembled = PETSC_FALSE;
625: if (roworiented) {
626: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
627: } else {
628: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
629: }
630: }
631: }
632: }
633: return(0);
634: }
636: /*
637: This function sets the j and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
638: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
639: No off-processor parts off the matrix are allowed here and mat->was_assembled has to be PETSC_FALSE.
640: */
641: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat_Symbolic(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[])
642: {
643: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
644: Mat A = aij->A; /* diagonal part of the matrix */
645: Mat B = aij->B; /* offdiagonal part of the matrix */
646: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
647: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
648: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,col;
649: PetscInt *ailen = a->ilen,*aj = a->j;
650: PetscInt *bilen = b->ilen,*bj = b->j;
651: PetscInt am = aij->A->rmap->n,j;
652: PetscInt diag_so_far = 0,dnz;
653: PetscInt offd_so_far = 0,onz;
656: /* Iterate over all rows of the matrix */
657: for (j=0; j<am; j++) {
658: dnz = onz = 0;
659: /* Iterate over all non-zero columns of the current row */
660: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
661: /* If column is in the diagonal */
662: if (mat_j[col] >= cstart && mat_j[col] < cend) {
663: aj[diag_so_far++] = mat_j[col] - cstart;
664: dnz++;
665: } else { /* off-diagonal entries */
666: bj[offd_so_far++] = mat_j[col];
667: onz++;
668: }
669: }
670: ailen[j] = dnz;
671: bilen[j] = onz;
672: }
673: return(0);
674: }
676: /*
677: This function sets the local j, a and ilen arrays (of the diagonal and off-diagonal part) of an MPIAIJ-matrix.
678: The values in mat_i have to be sorted and the values in mat_j have to be sorted for each row (CSR-like).
679: No off-processor parts off the matrix are allowed here, they are set at a later point by MatSetValues_MPIAIJ.
680: Also, mat->was_assembled has to be false, otherwise the statement aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
681: would not be true and the more complex MatSetValues_MPIAIJ has to be used.
682: */
683: PetscErrorCode MatSetValues_MPIAIJ_CopyFromCSRFormat(Mat mat,const PetscInt mat_j[],const PetscInt mat_i[],const PetscScalar mat_a[])
684: {
685: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
686: Mat A = aij->A; /* diagonal part of the matrix */
687: Mat B = aij->B; /* offdiagonal part of the matrix */
688: Mat_SeqAIJ *aijd =(Mat_SeqAIJ*)(aij->A)->data,*aijo=(Mat_SeqAIJ*)(aij->B)->data;
689: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
690: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
691: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend;
692: PetscInt *ailen = a->ilen,*aj = a->j;
693: PetscInt *bilen = b->ilen,*bj = b->j;
694: PetscInt am = aij->A->rmap->n,j;
695: PetscInt *full_diag_i=aijd->i,*full_offd_i=aijo->i; /* These variables can also include non-local elements, which are set at a later point. */
696: PetscInt col,dnz_row,onz_row,rowstart_diag,rowstart_offd;
697: PetscScalar *aa = a->a,*ba = b->a;
700: /* Iterate over all rows of the matrix */
701: for (j=0; j<am; j++) {
702: dnz_row = onz_row = 0;
703: rowstart_offd = full_offd_i[j];
704: rowstart_diag = full_diag_i[j];
705: /* Iterate over all non-zero columns of the current row */
706: for (col=mat_i[j]; col<mat_i[j+1]; col++) {
707: /* If column is in the diagonal */
708: if (mat_j[col] >= cstart && mat_j[col] < cend) {
709: aj[rowstart_diag+dnz_row] = mat_j[col] - cstart;
710: aa[rowstart_diag+dnz_row] = mat_a[col];
711: dnz_row++;
712: } else { /* off-diagonal entries */
713: bj[rowstart_offd+onz_row] = mat_j[col];
714: ba[rowstart_offd+onz_row] = mat_a[col];
715: onz_row++;
716: }
717: }
718: ailen[j] = dnz_row;
719: bilen[j] = onz_row;
720: }
721: return(0);
722: }
724: PetscErrorCode MatGetValues_MPIAIJ(Mat mat,PetscInt m,const PetscInt idxm[],PetscInt n,const PetscInt idxn[],PetscScalar v[])
725: {
726: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
728: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
729: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
732: for (i=0; i<m; i++) {
733: if (idxm[i] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative row: %D",idxm[i]);*/
734: if (idxm[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",idxm[i],mat->rmap->N-1);
735: if (idxm[i] >= rstart && idxm[i] < rend) {
736: row = idxm[i] - rstart;
737: for (j=0; j<n; j++) {
738: if (idxn[j] < 0) continue; /* SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Negative column: %D",idxn[j]); */
739: if (idxn[j] >= mat->cmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",idxn[j],mat->cmap->N-1);
740: if (idxn[j] >= cstart && idxn[j] < cend) {
741: col = idxn[j] - cstart;
742: MatGetValues(aij->A,1,&row,1,&col,v+i*n+j);
743: } else {
744: if (!aij->colmap) {
745: MatCreateColmap_MPIAIJ_Private(mat);
746: }
747: #if defined(PETSC_USE_CTABLE)
748: PetscTableFind(aij->colmap,idxn[j]+1,&col);
749: col--;
750: #else
751: col = aij->colmap[idxn[j]] - 1;
752: #endif
753: if ((col < 0) || (aij->garray[col] != idxn[j])) *(v+i*n+j) = 0.0;
754: else {
755: MatGetValues(aij->B,1,&row,1,&col,v+i*n+j);
756: }
757: }
758: }
759: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only local values currently supported");
760: }
761: return(0);
762: }
764: extern PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat,Vec,Vec);
766: PetscErrorCode MatAssemblyBegin_MPIAIJ(Mat mat,MatAssemblyType mode)
767: {
768: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
770: PetscInt nstash,reallocs;
773: if (aij->donotstash || mat->nooffprocentries) return(0);
775: MatStashScatterBegin_Private(mat,&mat->stash,mat->rmap->range);
776: MatStashGetInfo_Private(&mat->stash,&nstash,&reallocs);
777: PetscInfo2(aij->A,"Stash has %D entries, uses %D mallocs.\n",nstash,reallocs);
778: return(0);
779: }
781: PetscErrorCode MatAssemblyEnd_MPIAIJ(Mat mat,MatAssemblyType mode)
782: {
783: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
784: Mat_SeqAIJ *a = (Mat_SeqAIJ*)aij->A->data;
786: PetscMPIInt n;
787: PetscInt i,j,rstart,ncols,flg;
788: PetscInt *row,*col;
789: PetscBool other_disassembled;
790: PetscScalar *val;
792: /* do not use 'b = (Mat_SeqAIJ*)aij->B->data' as B can be reset in disassembly */
795: if (!aij->donotstash && !mat->nooffprocentries) {
796: while (1) {
797: MatStashScatterGetMesg_Private(&mat->stash,&n,&row,&col,&val,&flg);
798: if (!flg) break;
800: for (i=0; i<n; ) {
801: /* Now identify the consecutive vals belonging to the same row */
802: for (j=i,rstart=row[j]; j<n; j++) {
803: if (row[j] != rstart) break;
804: }
805: if (j < n) ncols = j-i;
806: else ncols = n-i;
807: /* Now assemble all these values with a single function call */
808: MatSetValues_MPIAIJ(mat,1,row+i,ncols,col+i,val+i,mat->insertmode);
810: i = j;
811: }
812: }
813: MatStashScatterEnd_Private(&mat->stash);
814: }
815: MatAssemblyBegin(aij->A,mode);
816: MatAssemblyEnd(aij->A,mode);
818: /* determine if any processor has disassembled, if so we must
819: also disassemble ourselfs, in order that we may reassemble. */
820: /*
821: if nonzero structure of submatrix B cannot change then we know that
822: no processor disassembled thus we can skip this stuff
823: */
824: if (!((Mat_SeqAIJ*)aij->B->data)->nonew) {
825: MPIU_Allreduce(&mat->was_assembled,&other_disassembled,1,MPIU_BOOL,MPI_PROD,PetscObjectComm((PetscObject)mat));
826: if (mat->was_assembled && !other_disassembled) {
827: MatDisAssemble_MPIAIJ(mat);
828: }
829: }
830: if (!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) {
831: MatSetUpMultiply_MPIAIJ(mat);
832: }
833: MatSetOption(aij->B,MAT_USE_INODES,PETSC_FALSE);
834: MatAssemblyBegin(aij->B,mode);
835: MatAssemblyEnd(aij->B,mode);
837: PetscFree2(aij->rowvalues,aij->rowindices);
839: aij->rowvalues = 0;
841: VecDestroy(&aij->diag);
842: if (a->inode.size) mat->ops->multdiagonalblock = MatMultDiagonalBlock_MPIAIJ;
844: /* if no new nonzero locations are allowed in matrix then only set the matrix state the first time through */
845: if ((!mat->was_assembled && mode == MAT_FINAL_ASSEMBLY) || !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
846: PetscObjectState state = aij->A->nonzerostate + aij->B->nonzerostate;
847: MPIU_Allreduce(&state,&mat->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)mat));
848: }
849: return(0);
850: }
852: PetscErrorCode MatZeroEntries_MPIAIJ(Mat A)
853: {
854: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
858: MatZeroEntries(l->A);
859: MatZeroEntries(l->B);
860: return(0);
861: }
863: PetscErrorCode MatZeroRows_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
864: {
865: Mat_MPIAIJ *mat = (Mat_MPIAIJ *) A->data;
866: PetscObjectState sA, sB;
867: PetscInt *lrows;
868: PetscInt r, len;
869: PetscBool cong, lch, gch;
870: PetscErrorCode ierr;
873: /* get locally owned rows */
874: MatZeroRowsMapLocal_Private(A,N,rows,&len,&lrows);
875: MatHasCongruentLayouts(A,&cong);
876: /* fix right hand side if needed */
877: if (x && b) {
878: const PetscScalar *xx;
879: PetscScalar *bb;
881: if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
882: VecGetArrayRead(x, &xx);
883: VecGetArray(b, &bb);
884: for (r = 0; r < len; ++r) bb[lrows[r]] = diag*xx[lrows[r]];
885: VecRestoreArrayRead(x, &xx);
886: VecRestoreArray(b, &bb);
887: }
889: sA = mat->A->nonzerostate;
890: sB = mat->B->nonzerostate;
892: if (diag != 0.0 && cong) {
893: MatZeroRows(mat->A, len, lrows, diag, NULL, NULL);
894: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
895: } else if (diag != 0.0) { /* non-square or non congruent layouts -> if keepnonzeropattern is false, we allow for new insertion */
896: Mat_SeqAIJ *aijA = (Mat_SeqAIJ*)mat->A->data;
897: Mat_SeqAIJ *aijB = (Mat_SeqAIJ*)mat->B->data;
898: PetscInt nnwA, nnwB;
899: PetscBool nnzA, nnzB;
901: nnwA = aijA->nonew;
902: nnwB = aijB->nonew;
903: nnzA = aijA->keepnonzeropattern;
904: nnzB = aijB->keepnonzeropattern;
905: if (!nnzA) {
906: PetscInfo(mat->A,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on diagonal block.\n");
907: aijA->nonew = 0;
908: }
909: if (!nnzB) {
910: PetscInfo(mat->B,"Requested to not keep the pattern and add a nonzero diagonal; may encounter reallocations on off-diagonal block.\n");
911: aijB->nonew = 0;
912: }
913: /* Must zero here before the next loop */
914: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
915: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
916: for (r = 0; r < len; ++r) {
917: const PetscInt row = lrows[r] + A->rmap->rstart;
918: if (row >= A->cmap->N) continue;
919: MatSetValues(A, 1, &row, 1, &row, &diag, INSERT_VALUES);
920: }
921: aijA->nonew = nnwA;
922: aijB->nonew = nnwB;
923: } else {
924: MatZeroRows(mat->A, len, lrows, 0.0, NULL, NULL);
925: MatZeroRows(mat->B, len, lrows, 0.0, NULL, NULL);
926: }
927: PetscFree(lrows);
928: MatAssemblyBegin(A,MAT_FINAL_ASSEMBLY);
929: MatAssemblyEnd(A,MAT_FINAL_ASSEMBLY);
931: /* reduce nonzerostate */
932: lch = (PetscBool)(sA != mat->A->nonzerostate || sB != mat->B->nonzerostate);
933: MPIU_Allreduce(&lch,&gch,1,MPIU_BOOL,MPI_LOR,PetscObjectComm((PetscObject)A));
934: if (gch) A->nonzerostate++;
935: return(0);
936: }
938: PetscErrorCode MatZeroRowsColumns_MPIAIJ(Mat A,PetscInt N,const PetscInt rows[],PetscScalar diag,Vec x,Vec b)
939: {
940: Mat_MPIAIJ *l = (Mat_MPIAIJ*)A->data;
941: PetscErrorCode ierr;
942: PetscMPIInt n = A->rmap->n;
943: PetscInt i,j,r,m,p = 0,len = 0;
944: PetscInt *lrows,*owners = A->rmap->range;
945: PetscSFNode *rrows;
946: PetscSF sf;
947: const PetscScalar *xx;
948: PetscScalar *bb,*mask;
949: Vec xmask,lmask;
950: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)l->B->data;
951: const PetscInt *aj, *ii,*ridx;
952: PetscScalar *aa;
955: /* Create SF where leaves are input rows and roots are owned rows */
956: PetscMalloc1(n, &lrows);
957: for (r = 0; r < n; ++r) lrows[r] = -1;
958: PetscMalloc1(N, &rrows);
959: for (r = 0; r < N; ++r) {
960: const PetscInt idx = rows[r];
961: if (idx < 0 || A->rmap->N <= idx) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row %D out of range [0,%D)",idx,A->rmap->N);
962: if (idx < owners[p] || owners[p+1] <= idx) { /* short-circuit the search if the last p owns this row too */
963: PetscLayoutFindOwner(A->rmap,idx,&p);
964: }
965: rrows[r].rank = p;
966: rrows[r].index = rows[r] - owners[p];
967: }
968: PetscSFCreate(PetscObjectComm((PetscObject) A), &sf);
969: PetscSFSetGraph(sf, n, N, NULL, PETSC_OWN_POINTER, rrows, PETSC_OWN_POINTER);
970: /* Collect flags for rows to be zeroed */
971: PetscSFReduceBegin(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
972: PetscSFReduceEnd(sf, MPIU_INT, (PetscInt *) rows, lrows, MPI_LOR);
973: PetscSFDestroy(&sf);
974: /* Compress and put in row numbers */
975: for (r = 0; r < n; ++r) if (lrows[r] >= 0) lrows[len++] = r;
976: /* zero diagonal part of matrix */
977: MatZeroRowsColumns(l->A,len,lrows,diag,x,b);
978: /* handle off diagonal part of matrix */
979: MatCreateVecs(A,&xmask,NULL);
980: VecDuplicate(l->lvec,&lmask);
981: VecGetArray(xmask,&bb);
982: for (i=0; i<len; i++) bb[lrows[i]] = 1;
983: VecRestoreArray(xmask,&bb);
984: VecScatterBegin(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
985: VecScatterEnd(l->Mvctx,xmask,lmask,ADD_VALUES,SCATTER_FORWARD);
986: VecDestroy(&xmask);
987: if (x && b) { /* this code is buggy when the row and column layout don't match */
988: PetscBool cong;
990: MatHasCongruentLayouts(A,&cong);
991: if (!cong) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Need matching row/col layout");
992: VecScatterBegin(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
993: VecScatterEnd(l->Mvctx,x,l->lvec,INSERT_VALUES,SCATTER_FORWARD);
994: VecGetArrayRead(l->lvec,&xx);
995: VecGetArray(b,&bb);
996: }
997: VecGetArray(lmask,&mask);
998: /* remove zeroed rows of off diagonal matrix */
999: ii = aij->i;
1000: for (i=0; i<len; i++) {
1001: PetscMemzero(aij->a + ii[lrows[i]],(ii[lrows[i]+1] - ii[lrows[i]])*sizeof(PetscScalar));
1002: }
1003: /* loop over all elements of off process part of matrix zeroing removed columns*/
1004: if (aij->compressedrow.use) {
1005: m = aij->compressedrow.nrows;
1006: ii = aij->compressedrow.i;
1007: ridx = aij->compressedrow.rindex;
1008: for (i=0; i<m; i++) {
1009: n = ii[i+1] - ii[i];
1010: aj = aij->j + ii[i];
1011: aa = aij->a + ii[i];
1013: for (j=0; j<n; j++) {
1014: if (PetscAbsScalar(mask[*aj])) {
1015: if (b) bb[*ridx] -= *aa*xx[*aj];
1016: *aa = 0.0;
1017: }
1018: aa++;
1019: aj++;
1020: }
1021: ridx++;
1022: }
1023: } else { /* do not use compressed row format */
1024: m = l->B->rmap->n;
1025: for (i=0; i<m; i++) {
1026: n = ii[i+1] - ii[i];
1027: aj = aij->j + ii[i];
1028: aa = aij->a + ii[i];
1029: for (j=0; j<n; j++) {
1030: if (PetscAbsScalar(mask[*aj])) {
1031: if (b) bb[i] -= *aa*xx[*aj];
1032: *aa = 0.0;
1033: }
1034: aa++;
1035: aj++;
1036: }
1037: }
1038: }
1039: if (x && b) {
1040: VecRestoreArray(b,&bb);
1041: VecRestoreArrayRead(l->lvec,&xx);
1042: }
1043: VecRestoreArray(lmask,&mask);
1044: VecDestroy(&lmask);
1045: PetscFree(lrows);
1047: /* only change matrix nonzero state if pattern was allowed to be changed */
1048: if (!((Mat_SeqAIJ*)(l->A->data))->keepnonzeropattern) {
1049: PetscObjectState state = l->A->nonzerostate + l->B->nonzerostate;
1050: MPIU_Allreduce(&state,&A->nonzerostate,1,MPIU_INT64,MPI_SUM,PetscObjectComm((PetscObject)A));
1051: }
1052: return(0);
1053: }
1055: PetscErrorCode MatMult_MPIAIJ(Mat A,Vec xx,Vec yy)
1056: {
1057: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1059: PetscInt nt;
1060: VecScatter Mvctx = a->Mvctx;
1063: VecGetLocalSize(xx,&nt);
1064: if (nt != A->cmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Incompatible partition of A (%D) and xx (%D)",A->cmap->n,nt);
1066: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1067: (*a->A->ops->mult)(a->A,xx,yy);
1068: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1069: (*a->B->ops->multadd)(a->B,a->lvec,yy,yy);
1070: return(0);
1071: }
1073: PetscErrorCode MatMultDiagonalBlock_MPIAIJ(Mat A,Vec bb,Vec xx)
1074: {
1075: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1079: MatMultDiagonalBlock(a->A,bb,xx);
1080: return(0);
1081: }
1083: PetscErrorCode MatMultAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1084: {
1085: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1087: VecScatter Mvctx = a->Mvctx;
1090: if (a->Mvctx_mpi1_flg) Mvctx = a->Mvctx_mpi1;
1091: VecScatterBegin(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1092: (*a->A->ops->multadd)(a->A,xx,yy,zz);
1093: VecScatterEnd(Mvctx,xx,a->lvec,INSERT_VALUES,SCATTER_FORWARD);
1094: (*a->B->ops->multadd)(a->B,a->lvec,zz,zz);
1095: return(0);
1096: }
1098: PetscErrorCode MatMultTranspose_MPIAIJ(Mat A,Vec xx,Vec yy)
1099: {
1100: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1104: /* do nondiagonal part */
1105: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1106: /* do local part */
1107: (*a->A->ops->multtranspose)(a->A,xx,yy);
1108: /* add partial results together */
1109: VecScatterBegin(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1110: VecScatterEnd(a->Mvctx,a->lvec,yy,ADD_VALUES,SCATTER_REVERSE);
1111: return(0);
1112: }
1114: PetscErrorCode MatIsTranspose_MPIAIJ(Mat Amat,Mat Bmat,PetscReal tol,PetscBool *f)
1115: {
1116: MPI_Comm comm;
1117: Mat_MPIAIJ *Aij = (Mat_MPIAIJ*) Amat->data, *Bij;
1118: Mat Adia = Aij->A, Bdia, Aoff,Boff,*Aoffs,*Boffs;
1119: IS Me,Notme;
1121: PetscInt M,N,first,last,*notme,i;
1122: PetscBool lf;
1123: PetscMPIInt size;
1126: /* Easy test: symmetric diagonal block */
1127: Bij = (Mat_MPIAIJ*) Bmat->data; Bdia = Bij->A;
1128: MatIsTranspose(Adia,Bdia,tol,&lf);
1129: MPIU_Allreduce(&lf,f,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)Amat));
1130: if (!*f) return(0);
1131: PetscObjectGetComm((PetscObject)Amat,&comm);
1132: MPI_Comm_size(comm,&size);
1133: if (size == 1) return(0);
1135: /* Hard test: off-diagonal block. This takes a MatCreateSubMatrix. */
1136: MatGetSize(Amat,&M,&N);
1137: MatGetOwnershipRange(Amat,&first,&last);
1138: PetscMalloc1(N-last+first,¬me);
1139: for (i=0; i<first; i++) notme[i] = i;
1140: for (i=last; i<M; i++) notme[i-last+first] = i;
1141: ISCreateGeneral(MPI_COMM_SELF,N-last+first,notme,PETSC_COPY_VALUES,&Notme);
1142: ISCreateStride(MPI_COMM_SELF,last-first,first,1,&Me);
1143: MatCreateSubMatrices(Amat,1,&Me,&Notme,MAT_INITIAL_MATRIX,&Aoffs);
1144: Aoff = Aoffs[0];
1145: MatCreateSubMatrices(Bmat,1,&Notme,&Me,MAT_INITIAL_MATRIX,&Boffs);
1146: Boff = Boffs[0];
1147: MatIsTranspose(Aoff,Boff,tol,f);
1148: MatDestroyMatrices(1,&Aoffs);
1149: MatDestroyMatrices(1,&Boffs);
1150: ISDestroy(&Me);
1151: ISDestroy(&Notme);
1152: PetscFree(notme);
1153: return(0);
1154: }
1156: PetscErrorCode MatIsSymmetric_MPIAIJ(Mat A,PetscReal tol,PetscBool *f)
1157: {
1161: MatIsTranspose_MPIAIJ(A,A,tol,f);
1162: return(0);
1163: }
1165: PetscErrorCode MatMultTransposeAdd_MPIAIJ(Mat A,Vec xx,Vec yy,Vec zz)
1166: {
1167: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1171: /* do nondiagonal part */
1172: (*a->B->ops->multtranspose)(a->B,xx,a->lvec);
1173: /* do local part */
1174: (*a->A->ops->multtransposeadd)(a->A,xx,yy,zz);
1175: /* add partial results together */
1176: VecScatterBegin(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1177: VecScatterEnd(a->Mvctx,a->lvec,zz,ADD_VALUES,SCATTER_REVERSE);
1178: return(0);
1179: }
1181: /*
1182: This only works correctly for square matrices where the subblock A->A is the
1183: diagonal block
1184: */
1185: PetscErrorCode MatGetDiagonal_MPIAIJ(Mat A,Vec v)
1186: {
1188: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1191: if (A->rmap->N != A->cmap->N) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"Supports only square matrix where A->A is diag block");
1192: if (A->rmap->rstart != A->cmap->rstart || A->rmap->rend != A->cmap->rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"row partition must equal col partition");
1193: MatGetDiagonal(a->A,v);
1194: return(0);
1195: }
1197: PetscErrorCode MatScale_MPIAIJ(Mat A,PetscScalar aa)
1198: {
1199: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1203: MatScale(a->A,aa);
1204: MatScale(a->B,aa);
1205: return(0);
1206: }
1208: PetscErrorCode MatDestroy_MPIAIJ(Mat mat)
1209: {
1210: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1214: #if defined(PETSC_USE_LOG)
1215: PetscLogObjectState((PetscObject)mat,"Rows=%D, Cols=%D",mat->rmap->N,mat->cmap->N);
1216: #endif
1217: MatStashDestroy_Private(&mat->stash);
1218: VecDestroy(&aij->diag);
1219: MatDestroy(&aij->A);
1220: MatDestroy(&aij->B);
1221: #if defined(PETSC_USE_CTABLE)
1222: PetscTableDestroy(&aij->colmap);
1223: #else
1224: PetscFree(aij->colmap);
1225: #endif
1226: PetscFree(aij->garray);
1227: VecDestroy(&aij->lvec);
1228: VecScatterDestroy(&aij->Mvctx);
1229: if (aij->Mvctx_mpi1) {VecScatterDestroy(&aij->Mvctx_mpi1);}
1230: PetscFree2(aij->rowvalues,aij->rowindices);
1231: PetscFree(aij->ld);
1232: PetscFree(mat->data);
1234: PetscObjectChangeTypeName((PetscObject)mat,0);
1235: PetscObjectComposeFunction((PetscObject)mat,"MatStoreValues_C",NULL);
1236: PetscObjectComposeFunction((PetscObject)mat,"MatRetrieveValues_C",NULL);
1237: PetscObjectComposeFunction((PetscObject)mat,"MatIsTranspose_C",NULL);
1238: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocation_C",NULL);
1239: PetscObjectComposeFunction((PetscObject)mat,"MatResetPreallocation_C",NULL);
1240: PetscObjectComposeFunction((PetscObject)mat,"MatMPIAIJSetPreallocationCSR_C",NULL);
1241: PetscObjectComposeFunction((PetscObject)mat,"MatDiagonalScaleLocal_C",NULL);
1242: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_mpisbaij_C",NULL);
1243: #if defined(PETSC_HAVE_ELEMENTAL)
1244: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_elemental_C",NULL);
1245: #endif
1246: #if defined(PETSC_HAVE_HYPRE)
1247: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_hypre_C",NULL);
1248: PetscObjectComposeFunction((PetscObject)mat,"MatMatMatMult_transpose_mpiaij_mpiaij_C",NULL);
1249: #endif
1250: PetscObjectComposeFunction((PetscObject)mat,"MatConvert_mpiaij_is_C",NULL);
1251: PetscObjectComposeFunction((PetscObject)mat,"MatPtAP_is_mpiaij_C",NULL);
1252: return(0);
1253: }
1255: PetscErrorCode MatView_MPIAIJ_Binary(Mat mat,PetscViewer viewer)
1256: {
1257: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1258: Mat_SeqAIJ *A = (Mat_SeqAIJ*)aij->A->data;
1259: Mat_SeqAIJ *B = (Mat_SeqAIJ*)aij->B->data;
1261: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
1262: int fd;
1263: PetscInt nz,header[4],*row_lengths,*range=0,rlen,i;
1264: PetscInt nzmax,*column_indices,j,k,col,*garray = aij->garray,cnt,cstart = mat->cmap->rstart,rnz = 0;
1265: PetscScalar *column_values;
1266: PetscInt message_count,flowcontrolcount;
1267: FILE *file;
1270: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1271: MPI_Comm_size(PetscObjectComm((PetscObject)mat),&size);
1272: nz = A->nz + B->nz;
1273: PetscViewerBinaryGetDescriptor(viewer,&fd);
1274: if (!rank) {
1275: header[0] = MAT_FILE_CLASSID;
1276: header[1] = mat->rmap->N;
1277: header[2] = mat->cmap->N;
1279: MPI_Reduce(&nz,&header[3],1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1280: PetscBinaryWrite(fd,header,4,PETSC_INT,PETSC_TRUE);
1281: /* get largest number of rows any processor has */
1282: rlen = mat->rmap->n;
1283: range = mat->rmap->range;
1284: for (i=1; i<size; i++) rlen = PetscMax(rlen,range[i+1] - range[i]);
1285: } else {
1286: MPI_Reduce(&nz,0,1,MPIU_INT,MPI_SUM,0,PetscObjectComm((PetscObject)mat));
1287: rlen = mat->rmap->n;
1288: }
1290: /* load up the local row counts */
1291: PetscMalloc1(rlen+1,&row_lengths);
1292: for (i=0; i<mat->rmap->n; i++) row_lengths[i] = A->i[i+1] - A->i[i] + B->i[i+1] - B->i[i];
1294: /* store the row lengths to the file */
1295: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1296: if (!rank) {
1297: PetscBinaryWrite(fd,row_lengths,mat->rmap->n,PETSC_INT,PETSC_TRUE);
1298: for (i=1; i<size; i++) {
1299: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1300: rlen = range[i+1] - range[i];
1301: MPIULong_Recv(row_lengths,rlen,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1302: PetscBinaryWrite(fd,row_lengths,rlen,PETSC_INT,PETSC_TRUE);
1303: }
1304: PetscViewerFlowControlEndMaster(viewer,&message_count);
1305: } else {
1306: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1307: MPIULong_Send(row_lengths,mat->rmap->n,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1308: PetscViewerFlowControlEndWorker(viewer,&message_count);
1309: }
1310: PetscFree(row_lengths);
1312: /* load up the local column indices */
1313: nzmax = nz; /* th processor needs space a largest processor needs */
1314: MPI_Reduce(&nz,&nzmax,1,MPIU_INT,MPI_MAX,0,PetscObjectComm((PetscObject)mat));
1315: PetscMalloc1(nzmax+1,&column_indices);
1316: cnt = 0;
1317: for (i=0; i<mat->rmap->n; i++) {
1318: for (j=B->i[i]; j<B->i[i+1]; j++) {
1319: if ((col = garray[B->j[j]]) > cstart) break;
1320: column_indices[cnt++] = col;
1321: }
1322: for (k=A->i[i]; k<A->i[i+1]; k++) column_indices[cnt++] = A->j[k] + cstart;
1323: for (; j<B->i[i+1]; j++) column_indices[cnt++] = garray[B->j[j]];
1324: }
1325: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
1327: /* store the column indices to the file */
1328: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1329: if (!rank) {
1330: MPI_Status status;
1331: PetscBinaryWrite(fd,column_indices,nz,PETSC_INT,PETSC_TRUE);
1332: for (i=1; i<size; i++) {
1333: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1334: MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1335: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1336: MPIULong_Recv(column_indices,rnz,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat));
1337: PetscBinaryWrite(fd,column_indices,rnz,PETSC_INT,PETSC_TRUE);
1338: }
1339: PetscViewerFlowControlEndMaster(viewer,&message_count);
1340: } else {
1341: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1342: MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1343: MPIULong_Send(column_indices,nz,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1344: PetscViewerFlowControlEndWorker(viewer,&message_count);
1345: }
1346: PetscFree(column_indices);
1348: /* load up the local column values */
1349: PetscMalloc1(nzmax+1,&column_values);
1350: cnt = 0;
1351: for (i=0; i<mat->rmap->n; i++) {
1352: for (j=B->i[i]; j<B->i[i+1]; j++) {
1353: if (garray[B->j[j]] > cstart) break;
1354: column_values[cnt++] = B->a[j];
1355: }
1356: for (k=A->i[i]; k<A->i[i+1]; k++) column_values[cnt++] = A->a[k];
1357: for (; j<B->i[i+1]; j++) column_values[cnt++] = B->a[j];
1358: }
1359: if (cnt != A->nz + B->nz) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_PLIB,"Internal PETSc error: cnt = %D nz = %D",cnt,A->nz+B->nz);
1361: /* store the column values to the file */
1362: PetscViewerFlowControlStart(viewer,&message_count,&flowcontrolcount);
1363: if (!rank) {
1364: MPI_Status status;
1365: PetscBinaryWrite(fd,column_values,nz,PETSC_SCALAR,PETSC_TRUE);
1366: for (i=1; i<size; i++) {
1367: PetscViewerFlowControlStepMaster(viewer,i,&message_count,flowcontrolcount);
1368: MPI_Recv(&rnz,1,MPIU_INT,i,tag,PetscObjectComm((PetscObject)mat),&status);
1369: if (rnz > nzmax) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_LIB,"Internal PETSc error: nz = %D nzmax = %D",nz,nzmax);
1370: MPIULong_Recv(column_values,rnz,MPIU_SCALAR,i,tag,PetscObjectComm((PetscObject)mat));
1371: PetscBinaryWrite(fd,column_values,rnz,PETSC_SCALAR,PETSC_TRUE);
1372: }
1373: PetscViewerFlowControlEndMaster(viewer,&message_count);
1374: } else {
1375: PetscViewerFlowControlStepWorker(viewer,rank,&message_count);
1376: MPI_Send(&nz,1,MPIU_INT,0,tag,PetscObjectComm((PetscObject)mat));
1377: MPIULong_Send(column_values,nz,MPIU_SCALAR,0,tag,PetscObjectComm((PetscObject)mat));
1378: PetscViewerFlowControlEndWorker(viewer,&message_count);
1379: }
1380: PetscFree(column_values);
1382: PetscViewerBinaryGetInfoPointer(viewer,&file);
1383: if (file) fprintf(file,"-matload_block_size %d\n",(int)PetscAbs(mat->rmap->bs));
1384: return(0);
1385: }
1387: #include <petscdraw.h>
1388: PetscErrorCode MatView_MPIAIJ_ASCIIorDraworSocket(Mat mat,PetscViewer viewer)
1389: {
1390: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1391: PetscErrorCode ierr;
1392: PetscMPIInt rank = aij->rank,size = aij->size;
1393: PetscBool isdraw,iascii,isbinary;
1394: PetscViewer sviewer;
1395: PetscViewerFormat format;
1398: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1399: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1400: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1401: if (iascii) {
1402: PetscViewerGetFormat(viewer,&format);
1403: if (format == PETSC_VIEWER_LOAD_BALANCE) {
1404: PetscInt i,nmax = 0,nmin = PETSC_MAX_INT,navg = 0,*nz,nzlocal = ((Mat_SeqAIJ*) (aij->A->data))->nz + ((Mat_SeqAIJ*) (aij->B->data))->nz;
1405: PetscMalloc1(size,&nz);
1406: MPI_Allgather(&nzlocal,1,MPIU_INT,nz,1,MPIU_INT,PetscObjectComm((PetscObject)mat));
1407: for (i=0; i<(PetscInt)size; i++) {
1408: nmax = PetscMax(nmax,nz[i]);
1409: nmin = PetscMin(nmin,nz[i]);
1410: navg += nz[i];
1411: }
1412: PetscFree(nz);
1413: navg = navg/size;
1414: PetscViewerASCIIPrintf(viewer,"Load Balance - Nonzeros: Min %D avg %D max %D\n",nmin,navg,nmax);
1415: return(0);
1416: }
1417: PetscViewerGetFormat(viewer,&format);
1418: if (format == PETSC_VIEWER_ASCII_INFO_DETAIL) {
1419: MatInfo info;
1420: PetscBool inodes;
1422: MPI_Comm_rank(PetscObjectComm((PetscObject)mat),&rank);
1423: MatGetInfo(mat,MAT_LOCAL,&info);
1424: MatInodeGetInodeSizes(aij->A,NULL,(PetscInt**)&inodes,NULL);
1425: PetscViewerASCIIPushSynchronized(viewer);
1426: if (!inodes) {
1427: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, not using I-node routines\n",
1428: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1429: } else {
1430: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] Local rows %D nz %D nz alloced %D mem %g, using I-node routines\n",
1431: rank,mat->rmap->n,(PetscInt)info.nz_used,(PetscInt)info.nz_allocated,(double)info.memory);
1432: }
1433: MatGetInfo(aij->A,MAT_LOCAL,&info);
1434: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] on-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1435: MatGetInfo(aij->B,MAT_LOCAL,&info);
1436: PetscViewerASCIISynchronizedPrintf(viewer,"[%d] off-diagonal part: nz %D \n",rank,(PetscInt)info.nz_used);
1437: PetscViewerFlush(viewer);
1438: PetscViewerASCIIPopSynchronized(viewer);
1439: PetscViewerASCIIPrintf(viewer,"Information on VecScatter used in matrix-vector product: \n");
1440: VecScatterView(aij->Mvctx,viewer);
1441: return(0);
1442: } else if (format == PETSC_VIEWER_ASCII_INFO) {
1443: PetscInt inodecount,inodelimit,*inodes;
1444: MatInodeGetInodeSizes(aij->A,&inodecount,&inodes,&inodelimit);
1445: if (inodes) {
1446: PetscViewerASCIIPrintf(viewer,"using I-node (on process 0) routines: found %D nodes, limit used is %D\n",inodecount,inodelimit);
1447: } else {
1448: PetscViewerASCIIPrintf(viewer,"not using I-node (on process 0) routines\n");
1449: }
1450: return(0);
1451: } else if (format == PETSC_VIEWER_ASCII_FACTOR_INFO) {
1452: return(0);
1453: }
1454: } else if (isbinary) {
1455: if (size == 1) {
1456: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1457: MatView(aij->A,viewer);
1458: } else {
1459: MatView_MPIAIJ_Binary(mat,viewer);
1460: }
1461: return(0);
1462: } else if (iascii && size == 1) {
1463: PetscObjectSetName((PetscObject)aij->A,((PetscObject)mat)->name);
1464: MatView(aij->A,viewer);
1465: return(0);
1466: } else if (isdraw) {
1467: PetscDraw draw;
1468: PetscBool isnull;
1469: PetscViewerDrawGetDraw(viewer,0,&draw);
1470: PetscDrawIsNull(draw,&isnull);
1471: if (isnull) return(0);
1472: }
1474: { /* assemble the entire matrix onto first processor */
1475: Mat A = NULL, Av;
1476: IS isrow,iscol;
1478: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1479: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1480: MatCreateSubMatrix(mat,isrow,iscol,MAT_INITIAL_MATRIX,&A);
1481: MatMPIAIJGetSeqAIJ(A,&Av,NULL,NULL);
1482: /* The commented code uses MatCreateSubMatrices instead */
1483: /*
1484: Mat *AA, A = NULL, Av;
1485: IS isrow,iscol;
1487: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->rmap->N : 0,0,1,&isrow);
1488: ISCreateStride(PetscObjectComm((PetscObject)mat),!rank ? mat->cmap->N : 0,0,1,&iscol);
1489: MatCreateSubMatrices(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,&AA);
1490: if (!rank) {
1491: PetscObjectReference((PetscObject)AA[0]);
1492: A = AA[0];
1493: Av = AA[0];
1494: }
1495: MatDestroySubMatrices(1,&AA);
1496: */
1497: ISDestroy(&iscol);
1498: ISDestroy(&isrow);
1499: /*
1500: Everyone has to call to draw the matrix since the graphics waits are
1501: synchronized across all processors that share the PetscDraw object
1502: */
1503: PetscViewerGetSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1504: if (!rank) {
1505: if (((PetscObject)mat)->name) {
1506: PetscObjectSetName((PetscObject)Av,((PetscObject)mat)->name);
1507: }
1508: MatView_SeqAIJ(Av,sviewer);
1509: }
1510: PetscViewerRestoreSubViewer(viewer,PETSC_COMM_SELF,&sviewer);
1511: PetscViewerFlush(viewer);
1512: MatDestroy(&A);
1513: }
1514: return(0);
1515: }
1517: PetscErrorCode MatView_MPIAIJ(Mat mat,PetscViewer viewer)
1518: {
1520: PetscBool iascii,isdraw,issocket,isbinary;
1523: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERASCII,&iascii);
1524: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERDRAW,&isdraw);
1525: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
1526: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERSOCKET,&issocket);
1527: if (iascii || isdraw || isbinary || issocket) {
1528: MatView_MPIAIJ_ASCIIorDraworSocket(mat,viewer);
1529: }
1530: return(0);
1531: }
1533: PetscErrorCode MatSOR_MPIAIJ(Mat matin,Vec bb,PetscReal omega,MatSORType flag,PetscReal fshift,PetscInt its,PetscInt lits,Vec xx)
1534: {
1535: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1537: Vec bb1 = 0;
1538: PetscBool hasop;
1541: if (flag == SOR_APPLY_UPPER) {
1542: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1543: return(0);
1544: }
1546: if (its > 1 || ~flag & SOR_ZERO_INITIAL_GUESS || flag & SOR_EISENSTAT) {
1547: VecDuplicate(bb,&bb1);
1548: }
1550: if ((flag & SOR_LOCAL_SYMMETRIC_SWEEP) == SOR_LOCAL_SYMMETRIC_SWEEP) {
1551: if (flag & SOR_ZERO_INITIAL_GUESS) {
1552: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1553: its--;
1554: }
1556: while (its--) {
1557: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1558: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1560: /* update rhs: bb1 = bb - B*x */
1561: VecScale(mat->lvec,-1.0);
1562: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1564: /* local sweep */
1565: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_SYMMETRIC_SWEEP,fshift,lits,1,xx);
1566: }
1567: } else if (flag & SOR_LOCAL_FORWARD_SWEEP) {
1568: if (flag & SOR_ZERO_INITIAL_GUESS) {
1569: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1570: its--;
1571: }
1572: while (its--) {
1573: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1574: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1576: /* update rhs: bb1 = bb - B*x */
1577: VecScale(mat->lvec,-1.0);
1578: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1580: /* local sweep */
1581: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_FORWARD_SWEEP,fshift,lits,1,xx);
1582: }
1583: } else if (flag & SOR_LOCAL_BACKWARD_SWEEP) {
1584: if (flag & SOR_ZERO_INITIAL_GUESS) {
1585: (*mat->A->ops->sor)(mat->A,bb,omega,flag,fshift,lits,1,xx);
1586: its--;
1587: }
1588: while (its--) {
1589: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1590: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1592: /* update rhs: bb1 = bb - B*x */
1593: VecScale(mat->lvec,-1.0);
1594: (*mat->B->ops->multadd)(mat->B,mat->lvec,bb,bb1);
1596: /* local sweep */
1597: (*mat->A->ops->sor)(mat->A,bb1,omega,SOR_BACKWARD_SWEEP,fshift,lits,1,xx);
1598: }
1599: } else if (flag & SOR_EISENSTAT) {
1600: Vec xx1;
1602: VecDuplicate(bb,&xx1);
1603: (*mat->A->ops->sor)(mat->A,bb,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_BACKWARD_SWEEP),fshift,lits,1,xx);
1605: VecScatterBegin(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1606: VecScatterEnd(mat->Mvctx,xx,mat->lvec,INSERT_VALUES,SCATTER_FORWARD);
1607: if (!mat->diag) {
1608: MatCreateVecs(matin,&mat->diag,NULL);
1609: MatGetDiagonal(matin,mat->diag);
1610: }
1611: MatHasOperation(matin,MATOP_MULT_DIAGONAL_BLOCK,&hasop);
1612: if (hasop) {
1613: MatMultDiagonalBlock(matin,xx,bb1);
1614: } else {
1615: VecPointwiseMult(bb1,mat->diag,xx);
1616: }
1617: VecAYPX(bb1,(omega-2.0)/omega,bb);
1619: MatMultAdd(mat->B,mat->lvec,bb1,bb1);
1621: /* local sweep */
1622: (*mat->A->ops->sor)(mat->A,bb1,omega,(MatSORType)(SOR_ZERO_INITIAL_GUESS | SOR_LOCAL_FORWARD_SWEEP),fshift,lits,1,xx1);
1623: VecAXPY(xx,1.0,xx1);
1624: VecDestroy(&xx1);
1625: } else SETERRQ(PetscObjectComm((PetscObject)matin),PETSC_ERR_SUP,"Parallel SOR not supported");
1627: VecDestroy(&bb1);
1629: matin->factorerrortype = mat->A->factorerrortype;
1630: return(0);
1631: }
1633: PetscErrorCode MatPermute_MPIAIJ(Mat A,IS rowp,IS colp,Mat *B)
1634: {
1635: Mat aA,aB,Aperm;
1636: const PetscInt *rwant,*cwant,*gcols,*ai,*bi,*aj,*bj;
1637: PetscScalar *aa,*ba;
1638: PetscInt i,j,m,n,ng,anz,bnz,*dnnz,*onnz,*tdnnz,*tonnz,*rdest,*cdest,*work,*gcdest;
1639: PetscSF rowsf,sf;
1640: IS parcolp = NULL;
1641: PetscBool done;
1645: MatGetLocalSize(A,&m,&n);
1646: ISGetIndices(rowp,&rwant);
1647: ISGetIndices(colp,&cwant);
1648: PetscMalloc3(PetscMax(m,n),&work,m,&rdest,n,&cdest);
1650: /* Invert row permutation to find out where my rows should go */
1651: PetscSFCreate(PetscObjectComm((PetscObject)A),&rowsf);
1652: PetscSFSetGraphLayout(rowsf,A->rmap,A->rmap->n,NULL,PETSC_OWN_POINTER,rwant);
1653: PetscSFSetFromOptions(rowsf);
1654: for (i=0; i<m; i++) work[i] = A->rmap->rstart + i;
1655: PetscSFReduceBegin(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1656: PetscSFReduceEnd(rowsf,MPIU_INT,work,rdest,MPIU_REPLACE);
1658: /* Invert column permutation to find out where my columns should go */
1659: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1660: PetscSFSetGraphLayout(sf,A->cmap,A->cmap->n,NULL,PETSC_OWN_POINTER,cwant);
1661: PetscSFSetFromOptions(sf);
1662: for (i=0; i<n; i++) work[i] = A->cmap->rstart + i;
1663: PetscSFReduceBegin(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1664: PetscSFReduceEnd(sf,MPIU_INT,work,cdest,MPIU_REPLACE);
1665: PetscSFDestroy(&sf);
1667: ISRestoreIndices(rowp,&rwant);
1668: ISRestoreIndices(colp,&cwant);
1669: MatMPIAIJGetSeqAIJ(A,&aA,&aB,&gcols);
1671: /* Find out where my gcols should go */
1672: MatGetSize(aB,NULL,&ng);
1673: PetscMalloc1(ng,&gcdest);
1674: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
1675: PetscSFSetGraphLayout(sf,A->cmap,ng,NULL,PETSC_OWN_POINTER,gcols);
1676: PetscSFSetFromOptions(sf);
1677: PetscSFBcastBegin(sf,MPIU_INT,cdest,gcdest);
1678: PetscSFBcastEnd(sf,MPIU_INT,cdest,gcdest);
1679: PetscSFDestroy(&sf);
1681: PetscCalloc4(m,&dnnz,m,&onnz,m,&tdnnz,m,&tonnz);
1682: MatGetRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1683: MatGetRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1684: for (i=0; i<m; i++) {
1685: PetscInt row = rdest[i],rowner;
1686: PetscLayoutFindOwner(A->rmap,row,&rowner);
1687: for (j=ai[i]; j<ai[i+1]; j++) {
1688: PetscInt cowner,col = cdest[aj[j]];
1689: PetscLayoutFindOwner(A->cmap,col,&cowner); /* Could build an index for the columns to eliminate this search */
1690: if (rowner == cowner) dnnz[i]++;
1691: else onnz[i]++;
1692: }
1693: for (j=bi[i]; j<bi[i+1]; j++) {
1694: PetscInt cowner,col = gcdest[bj[j]];
1695: PetscLayoutFindOwner(A->cmap,col,&cowner);
1696: if (rowner == cowner) dnnz[i]++;
1697: else onnz[i]++;
1698: }
1699: }
1700: PetscSFBcastBegin(rowsf,MPIU_INT,dnnz,tdnnz);
1701: PetscSFBcastEnd(rowsf,MPIU_INT,dnnz,tdnnz);
1702: PetscSFBcastBegin(rowsf,MPIU_INT,onnz,tonnz);
1703: PetscSFBcastEnd(rowsf,MPIU_INT,onnz,tonnz);
1704: PetscSFDestroy(&rowsf);
1706: MatCreateAIJ(PetscObjectComm((PetscObject)A),A->rmap->n,A->cmap->n,A->rmap->N,A->cmap->N,0,tdnnz,0,tonnz,&Aperm);
1707: MatSeqAIJGetArray(aA,&aa);
1708: MatSeqAIJGetArray(aB,&ba);
1709: for (i=0; i<m; i++) {
1710: PetscInt *acols = dnnz,*bcols = onnz; /* Repurpose now-unneeded arrays */
1711: PetscInt j0,rowlen;
1712: rowlen = ai[i+1] - ai[i];
1713: for (j0=j=0; j<rowlen; j0=j) { /* rowlen could be larger than number of rows m, so sum in batches */
1714: for ( ; j<PetscMin(rowlen,j0+m); j++) acols[j-j0] = cdest[aj[ai[i]+j]];
1715: MatSetValues(Aperm,1,&rdest[i],j-j0,acols,aa+ai[i]+j0,INSERT_VALUES);
1716: }
1717: rowlen = bi[i+1] - bi[i];
1718: for (j0=j=0; j<rowlen; j0=j) {
1719: for ( ; j<PetscMin(rowlen,j0+m); j++) bcols[j-j0] = gcdest[bj[bi[i]+j]];
1720: MatSetValues(Aperm,1,&rdest[i],j-j0,bcols,ba+bi[i]+j0,INSERT_VALUES);
1721: }
1722: }
1723: MatAssemblyBegin(Aperm,MAT_FINAL_ASSEMBLY);
1724: MatAssemblyEnd(Aperm,MAT_FINAL_ASSEMBLY);
1725: MatRestoreRowIJ(aA,0,PETSC_FALSE,PETSC_FALSE,&anz,&ai,&aj,&done);
1726: MatRestoreRowIJ(aB,0,PETSC_FALSE,PETSC_FALSE,&bnz,&bi,&bj,&done);
1727: MatSeqAIJRestoreArray(aA,&aa);
1728: MatSeqAIJRestoreArray(aB,&ba);
1729: PetscFree4(dnnz,onnz,tdnnz,tonnz);
1730: PetscFree3(work,rdest,cdest);
1731: PetscFree(gcdest);
1732: if (parcolp) {ISDestroy(&colp);}
1733: *B = Aperm;
1734: return(0);
1735: }
1737: PetscErrorCode MatGetGhosts_MPIAIJ(Mat mat,PetscInt *nghosts,const PetscInt *ghosts[])
1738: {
1739: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1743: MatGetSize(aij->B,NULL,nghosts);
1744: if (ghosts) *ghosts = aij->garray;
1745: return(0);
1746: }
1748: PetscErrorCode MatGetInfo_MPIAIJ(Mat matin,MatInfoType flag,MatInfo *info)
1749: {
1750: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1751: Mat A = mat->A,B = mat->B;
1753: PetscReal isend[5],irecv[5];
1756: info->block_size = 1.0;
1757: MatGetInfo(A,MAT_LOCAL,info);
1759: isend[0] = info->nz_used; isend[1] = info->nz_allocated; isend[2] = info->nz_unneeded;
1760: isend[3] = info->memory; isend[4] = info->mallocs;
1762: MatGetInfo(B,MAT_LOCAL,info);
1764: isend[0] += info->nz_used; isend[1] += info->nz_allocated; isend[2] += info->nz_unneeded;
1765: isend[3] += info->memory; isend[4] += info->mallocs;
1766: if (flag == MAT_LOCAL) {
1767: info->nz_used = isend[0];
1768: info->nz_allocated = isend[1];
1769: info->nz_unneeded = isend[2];
1770: info->memory = isend[3];
1771: info->mallocs = isend[4];
1772: } else if (flag == MAT_GLOBAL_MAX) {
1773: MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)matin));
1775: info->nz_used = irecv[0];
1776: info->nz_allocated = irecv[1];
1777: info->nz_unneeded = irecv[2];
1778: info->memory = irecv[3];
1779: info->mallocs = irecv[4];
1780: } else if (flag == MAT_GLOBAL_SUM) {
1781: MPIU_Allreduce(isend,irecv,5,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)matin));
1783: info->nz_used = irecv[0];
1784: info->nz_allocated = irecv[1];
1785: info->nz_unneeded = irecv[2];
1786: info->memory = irecv[3];
1787: info->mallocs = irecv[4];
1788: }
1789: info->fill_ratio_given = 0; /* no parallel LU/ILU/Cholesky */
1790: info->fill_ratio_needed = 0;
1791: info->factor_mallocs = 0;
1792: return(0);
1793: }
1795: PetscErrorCode MatSetOption_MPIAIJ(Mat A,MatOption op,PetscBool flg)
1796: {
1797: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
1801: switch (op) {
1802: case MAT_NEW_NONZERO_LOCATIONS:
1803: case MAT_NEW_NONZERO_ALLOCATION_ERR:
1804: case MAT_UNUSED_NONZERO_LOCATION_ERR:
1805: case MAT_KEEP_NONZERO_PATTERN:
1806: case MAT_NEW_NONZERO_LOCATION_ERR:
1807: case MAT_USE_INODES:
1808: case MAT_IGNORE_ZERO_ENTRIES:
1809: MatCheckPreallocated(A,1);
1810: MatSetOption(a->A,op,flg);
1811: MatSetOption(a->B,op,flg);
1812: break;
1813: case MAT_ROW_ORIENTED:
1814: MatCheckPreallocated(A,1);
1815: a->roworiented = flg;
1817: MatSetOption(a->A,op,flg);
1818: MatSetOption(a->B,op,flg);
1819: break;
1820: case MAT_NEW_DIAGONALS:
1821: PetscInfo1(A,"Option %s ignored\n",MatOptions[op]);
1822: break;
1823: case MAT_IGNORE_OFF_PROC_ENTRIES:
1824: a->donotstash = flg;
1825: break;
1826: /* Symmetry flags are handled directly by MatSetOption() and they don't affect preallocation */
1827: case MAT_SPD:
1828: case MAT_SYMMETRIC:
1829: case MAT_STRUCTURALLY_SYMMETRIC:
1830: case MAT_HERMITIAN:
1831: case MAT_SYMMETRY_ETERNAL:
1832: break;
1833: case MAT_SUBMAT_SINGLEIS:
1834: A->submat_singleis = flg;
1835: break;
1836: case MAT_STRUCTURE_ONLY:
1837: /* The option is handled directly by MatSetOption() */
1838: break;
1839: default:
1840: SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_SUP,"unknown option %d",op);
1841: }
1842: return(0);
1843: }
1845: PetscErrorCode MatGetRow_MPIAIJ(Mat matin,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1846: {
1847: Mat_MPIAIJ *mat = (Mat_MPIAIJ*)matin->data;
1848: PetscScalar *vworkA,*vworkB,**pvA,**pvB,*v_p;
1850: PetscInt i,*cworkA,*cworkB,**pcA,**pcB,cstart = matin->cmap->rstart;
1851: PetscInt nztot,nzA,nzB,lrow,rstart = matin->rmap->rstart,rend = matin->rmap->rend;
1852: PetscInt *cmap,*idx_p;
1855: if (mat->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Already active");
1856: mat->getrowactive = PETSC_TRUE;
1858: if (!mat->rowvalues && (idx || v)) {
1859: /*
1860: allocate enough space to hold information from the longest row.
1861: */
1862: Mat_SeqAIJ *Aa = (Mat_SeqAIJ*)mat->A->data,*Ba = (Mat_SeqAIJ*)mat->B->data;
1863: PetscInt max = 1,tmp;
1864: for (i=0; i<matin->rmap->n; i++) {
1865: tmp = Aa->i[i+1] - Aa->i[i] + Ba->i[i+1] - Ba->i[i];
1866: if (max < tmp) max = tmp;
1867: }
1868: PetscMalloc2(max,&mat->rowvalues,max,&mat->rowindices);
1869: }
1871: if (row < rstart || row >= rend) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Only local rows");
1872: lrow = row - rstart;
1874: pvA = &vworkA; pcA = &cworkA; pvB = &vworkB; pcB = &cworkB;
1875: if (!v) {pvA = 0; pvB = 0;}
1876: if (!idx) {pcA = 0; if (!v) pcB = 0;}
1877: (*mat->A->ops->getrow)(mat->A,lrow,&nzA,pcA,pvA);
1878: (*mat->B->ops->getrow)(mat->B,lrow,&nzB,pcB,pvB);
1879: nztot = nzA + nzB;
1881: cmap = mat->garray;
1882: if (v || idx) {
1883: if (nztot) {
1884: /* Sort by increasing column numbers, assuming A and B already sorted */
1885: PetscInt imark = -1;
1886: if (v) {
1887: *v = v_p = mat->rowvalues;
1888: for (i=0; i<nzB; i++) {
1889: if (cmap[cworkB[i]] < cstart) v_p[i] = vworkB[i];
1890: else break;
1891: }
1892: imark = i;
1893: for (i=0; i<nzA; i++) v_p[imark+i] = vworkA[i];
1894: for (i=imark; i<nzB; i++) v_p[nzA+i] = vworkB[i];
1895: }
1896: if (idx) {
1897: *idx = idx_p = mat->rowindices;
1898: if (imark > -1) {
1899: for (i=0; i<imark; i++) {
1900: idx_p[i] = cmap[cworkB[i]];
1901: }
1902: } else {
1903: for (i=0; i<nzB; i++) {
1904: if (cmap[cworkB[i]] < cstart) idx_p[i] = cmap[cworkB[i]];
1905: else break;
1906: }
1907: imark = i;
1908: }
1909: for (i=0; i<nzA; i++) idx_p[imark+i] = cstart + cworkA[i];
1910: for (i=imark; i<nzB; i++) idx_p[nzA+i] = cmap[cworkB[i]];
1911: }
1912: } else {
1913: if (idx) *idx = 0;
1914: if (v) *v = 0;
1915: }
1916: }
1917: *nz = nztot;
1918: (*mat->A->ops->restorerow)(mat->A,lrow,&nzA,pcA,pvA);
1919: (*mat->B->ops->restorerow)(mat->B,lrow,&nzB,pcB,pvB);
1920: return(0);
1921: }
1923: PetscErrorCode MatRestoreRow_MPIAIJ(Mat mat,PetscInt row,PetscInt *nz,PetscInt **idx,PetscScalar **v)
1924: {
1925: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1928: if (!aij->getrowactive) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"MatGetRow() must be called first");
1929: aij->getrowactive = PETSC_FALSE;
1930: return(0);
1931: }
1933: PetscErrorCode MatNorm_MPIAIJ(Mat mat,NormType type,PetscReal *norm)
1934: {
1935: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
1936: Mat_SeqAIJ *amat = (Mat_SeqAIJ*)aij->A->data,*bmat = (Mat_SeqAIJ*)aij->B->data;
1938: PetscInt i,j,cstart = mat->cmap->rstart;
1939: PetscReal sum = 0.0;
1940: MatScalar *v;
1943: if (aij->size == 1) {
1944: MatNorm(aij->A,type,norm);
1945: } else {
1946: if (type == NORM_FROBENIUS) {
1947: v = amat->a;
1948: for (i=0; i<amat->nz; i++) {
1949: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1950: }
1951: v = bmat->a;
1952: for (i=0; i<bmat->nz; i++) {
1953: sum += PetscRealPart(PetscConj(*v)*(*v)); v++;
1954: }
1955: MPIU_Allreduce(&sum,norm,1,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1956: *norm = PetscSqrtReal(*norm);
1957: PetscLogFlops(2*amat->nz+2*bmat->nz);
1958: } else if (type == NORM_1) { /* max column norm */
1959: PetscReal *tmp,*tmp2;
1960: PetscInt *jj,*garray = aij->garray;
1961: PetscCalloc1(mat->cmap->N+1,&tmp);
1962: PetscMalloc1(mat->cmap->N+1,&tmp2);
1963: *norm = 0.0;
1964: v = amat->a; jj = amat->j;
1965: for (j=0; j<amat->nz; j++) {
1966: tmp[cstart + *jj++] += PetscAbsScalar(*v); v++;
1967: }
1968: v = bmat->a; jj = bmat->j;
1969: for (j=0; j<bmat->nz; j++) {
1970: tmp[garray[*jj++]] += PetscAbsScalar(*v); v++;
1971: }
1972: MPIU_Allreduce(tmp,tmp2,mat->cmap->N,MPIU_REAL,MPIU_SUM,PetscObjectComm((PetscObject)mat));
1973: for (j=0; j<mat->cmap->N; j++) {
1974: if (tmp2[j] > *norm) *norm = tmp2[j];
1975: }
1976: PetscFree(tmp);
1977: PetscFree(tmp2);
1978: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1979: } else if (type == NORM_INFINITY) { /* max row norm */
1980: PetscReal ntemp = 0.0;
1981: for (j=0; j<aij->A->rmap->n; j++) {
1982: v = amat->a + amat->i[j];
1983: sum = 0.0;
1984: for (i=0; i<amat->i[j+1]-amat->i[j]; i++) {
1985: sum += PetscAbsScalar(*v); v++;
1986: }
1987: v = bmat->a + bmat->i[j];
1988: for (i=0; i<bmat->i[j+1]-bmat->i[j]; i++) {
1989: sum += PetscAbsScalar(*v); v++;
1990: }
1991: if (sum > ntemp) ntemp = sum;
1992: }
1993: MPIU_Allreduce(&ntemp,norm,1,MPIU_REAL,MPIU_MAX,PetscObjectComm((PetscObject)mat));
1994: PetscLogFlops(PetscMax(amat->nz+bmat->nz-1,0));
1995: } else SETERRQ(PetscObjectComm((PetscObject)mat),PETSC_ERR_SUP,"No support for two norm");
1996: }
1997: return(0);
1998: }
2000: PetscErrorCode MatTranspose_MPIAIJ(Mat A,MatReuse reuse,Mat *matout)
2001: {
2002: Mat_MPIAIJ *a =(Mat_MPIAIJ*)A->data,*b;
2003: Mat_SeqAIJ *Aloc =(Mat_SeqAIJ*)a->A->data,*Bloc=(Mat_SeqAIJ*)a->B->data,*sub_B_diag;
2004: PetscInt M = A->rmap->N,N=A->cmap->N,ma,na,mb,nb,*ai,*aj,*bi,*bj,row,*cols,*cols_tmp,*B_diag_ilen,*B_diag_i,i,ncol,A_diag_ncol;
2006: Mat B,A_diag,*B_diag;
2007: MatScalar *array;
2010: ma = A->rmap->n; na = A->cmap->n; mb = a->B->rmap->n; nb = a->B->cmap->n;
2011: ai = Aloc->i; aj = Aloc->j;
2012: bi = Bloc->i; bj = Bloc->j;
2013: if (reuse == MAT_INITIAL_MATRIX || *matout == A) {
2014: PetscInt *d_nnz,*g_nnz,*o_nnz;
2015: PetscSFNode *oloc;
2016: PETSC_UNUSED PetscSF sf;
2018: PetscMalloc4(na,&d_nnz,na,&o_nnz,nb,&g_nnz,nb,&oloc);
2019: /* compute d_nnz for preallocation */
2020: PetscMemzero(d_nnz,na*sizeof(PetscInt));
2021: for (i=0; i<ai[ma]; i++) {
2022: d_nnz[aj[i]]++;
2023: }
2024: /* compute local off-diagonal contributions */
2025: PetscMemzero(g_nnz,nb*sizeof(PetscInt));
2026: for (i=0; i<bi[ma]; i++) g_nnz[bj[i]]++;
2027: /* map those to global */
2028: PetscSFCreate(PetscObjectComm((PetscObject)A),&sf);
2029: PetscSFSetGraphLayout(sf,A->cmap,nb,NULL,PETSC_USE_POINTER,a->garray);
2030: PetscSFSetFromOptions(sf);
2031: PetscMemzero(o_nnz,na*sizeof(PetscInt));
2032: PetscSFReduceBegin(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2033: PetscSFReduceEnd(sf,MPIU_INT,g_nnz,o_nnz,MPIU_SUM);
2034: PetscSFDestroy(&sf);
2036: MatCreate(PetscObjectComm((PetscObject)A),&B);
2037: MatSetSizes(B,A->cmap->n,A->rmap->n,N,M);
2038: MatSetBlockSizes(B,PetscAbs(A->cmap->bs),PetscAbs(A->rmap->bs));
2039: MatSetType(B,((PetscObject)A)->type_name);
2040: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
2041: PetscFree4(d_nnz,o_nnz,g_nnz,oloc);
2042: } else {
2043: B = *matout;
2044: MatSetOption(B,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_TRUE);
2045: }
2047: b = (Mat_MPIAIJ*)B->data;
2048: A_diag = a->A;
2049: B_diag = &b->A;
2050: sub_B_diag = (Mat_SeqAIJ*)(*B_diag)->data;
2051: A_diag_ncol = A_diag->cmap->N;
2052: B_diag_ilen = sub_B_diag->ilen;
2053: B_diag_i = sub_B_diag->i;
2055: /* Set ilen for diagonal of B */
2056: for (i=0; i<A_diag_ncol; i++) {
2057: B_diag_ilen[i] = B_diag_i[i+1] - B_diag_i[i];
2058: }
2060: /* Transpose the diagonal part of the matrix. In contrast to the offdiagonal part, this can be done
2061: very quickly (=without using MatSetValues), because all writes are local. */
2062: MatTranspose(A_diag,MAT_REUSE_MATRIX,B_diag);
2064: /* copy over the B part */
2065: PetscCalloc1(bi[mb],&cols);
2066: array = Bloc->a;
2067: row = A->rmap->rstart;
2068: for (i=0; i<bi[mb]; i++) cols[i] = a->garray[bj[i]];
2069: cols_tmp = cols;
2070: for (i=0; i<mb; i++) {
2071: ncol = bi[i+1]-bi[i];
2072: MatSetValues(B,ncol,cols_tmp,1,&row,array,INSERT_VALUES);
2073: row++;
2074: array += ncol; cols_tmp += ncol;
2075: }
2076: PetscFree(cols);
2078: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
2079: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
2080: if (reuse == MAT_INITIAL_MATRIX || reuse == MAT_REUSE_MATRIX) {
2081: *matout = B;
2082: } else {
2083: MatHeaderMerge(A,&B);
2084: }
2085: return(0);
2086: }
2088: PetscErrorCode MatDiagonalScale_MPIAIJ(Mat mat,Vec ll,Vec rr)
2089: {
2090: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2091: Mat a = aij->A,b = aij->B;
2093: PetscInt s1,s2,s3;
2096: MatGetLocalSize(mat,&s2,&s3);
2097: if (rr) {
2098: VecGetLocalSize(rr,&s1);
2099: if (s1!=s3) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"right vector non-conforming local size");
2100: /* Overlap communication with computation. */
2101: VecScatterBegin(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2102: }
2103: if (ll) {
2104: VecGetLocalSize(ll,&s1);
2105: if (s1!=s2) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"left vector non-conforming local size");
2106: (*b->ops->diagonalscale)(b,ll,0);
2107: }
2108: /* scale the diagonal block */
2109: (*a->ops->diagonalscale)(a,ll,rr);
2111: if (rr) {
2112: /* Do a scatter end and then right scale the off-diagonal block */
2113: VecScatterEnd(aij->Mvctx,rr,aij->lvec,INSERT_VALUES,SCATTER_FORWARD);
2114: (*b->ops->diagonalscale)(b,0,aij->lvec);
2115: }
2116: return(0);
2117: }
2119: PetscErrorCode MatSetUnfactored_MPIAIJ(Mat A)
2120: {
2121: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2125: MatSetUnfactored(a->A);
2126: return(0);
2127: }
2129: PetscErrorCode MatEqual_MPIAIJ(Mat A,Mat B,PetscBool *flag)
2130: {
2131: Mat_MPIAIJ *matB = (Mat_MPIAIJ*)B->data,*matA = (Mat_MPIAIJ*)A->data;
2132: Mat a,b,c,d;
2133: PetscBool flg;
2137: a = matA->A; b = matA->B;
2138: c = matB->A; d = matB->B;
2140: MatEqual(a,c,&flg);
2141: if (flg) {
2142: MatEqual(b,d,&flg);
2143: }
2144: MPIU_Allreduce(&flg,flag,1,MPIU_BOOL,MPI_LAND,PetscObjectComm((PetscObject)A));
2145: return(0);
2146: }
2148: PetscErrorCode MatCopy_MPIAIJ(Mat A,Mat B,MatStructure str)
2149: {
2151: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2152: Mat_MPIAIJ *b = (Mat_MPIAIJ*)B->data;
2155: /* If the two matrices don't have the same copy implementation, they aren't compatible for fast copy. */
2156: if ((str != SAME_NONZERO_PATTERN) || (A->ops->copy != B->ops->copy)) {
2157: /* because of the column compression in the off-processor part of the matrix a->B,
2158: the number of columns in a->B and b->B may be different, hence we cannot call
2159: the MatCopy() directly on the two parts. If need be, we can provide a more
2160: efficient copy than the MatCopy_Basic() by first uncompressing the a->B matrices
2161: then copying the submatrices */
2162: MatCopy_Basic(A,B,str);
2163: } else {
2164: MatCopy(a->A,b->A,str);
2165: MatCopy(a->B,b->B,str);
2166: }
2167: PetscObjectStateIncrease((PetscObject)B);
2168: return(0);
2169: }
2171: PetscErrorCode MatSetUp_MPIAIJ(Mat A)
2172: {
2176: MatMPIAIJSetPreallocation(A,PETSC_DEFAULT,0,PETSC_DEFAULT,0);
2177: return(0);
2178: }
2180: /*
2181: Computes the number of nonzeros per row needed for preallocation when X and Y
2182: have different nonzero structure.
2183: */
2184: PetscErrorCode MatAXPYGetPreallocation_MPIX_private(PetscInt m,const PetscInt *xi,const PetscInt *xj,const PetscInt *xltog,const PetscInt *yi,const PetscInt *yj,const PetscInt *yltog,PetscInt *nnz)
2185: {
2186: PetscInt i,j,k,nzx,nzy;
2189: /* Set the number of nonzeros in the new matrix */
2190: for (i=0; i<m; i++) {
2191: const PetscInt *xjj = xj+xi[i],*yjj = yj+yi[i];
2192: nzx = xi[i+1] - xi[i];
2193: nzy = yi[i+1] - yi[i];
2194: nnz[i] = 0;
2195: for (j=0,k=0; j<nzx; j++) { /* Point in X */
2196: for (; k<nzy && yltog[yjj[k]]<xltog[xjj[j]]; k++) nnz[i]++; /* Catch up to X */
2197: if (k<nzy && yltog[yjj[k]]==xltog[xjj[j]]) k++; /* Skip duplicate */
2198: nnz[i]++;
2199: }
2200: for (; k<nzy; k++) nnz[i]++;
2201: }
2202: return(0);
2203: }
2205: /* This is the same as MatAXPYGetPreallocation_SeqAIJ, except that the local-to-global map is provided */
2206: static PetscErrorCode MatAXPYGetPreallocation_MPIAIJ(Mat Y,const PetscInt *yltog,Mat X,const PetscInt *xltog,PetscInt *nnz)
2207: {
2209: PetscInt m = Y->rmap->N;
2210: Mat_SeqAIJ *x = (Mat_SeqAIJ*)X->data;
2211: Mat_SeqAIJ *y = (Mat_SeqAIJ*)Y->data;
2214: MatAXPYGetPreallocation_MPIX_private(m,x->i,x->j,xltog,y->i,y->j,yltog,nnz);
2215: return(0);
2216: }
2218: PetscErrorCode MatAXPY_MPIAIJ(Mat Y,PetscScalar a,Mat X,MatStructure str)
2219: {
2221: Mat_MPIAIJ *xx = (Mat_MPIAIJ*)X->data,*yy = (Mat_MPIAIJ*)Y->data;
2222: PetscBLASInt bnz,one=1;
2223: Mat_SeqAIJ *x,*y;
2226: if (str == SAME_NONZERO_PATTERN) {
2227: PetscScalar alpha = a;
2228: x = (Mat_SeqAIJ*)xx->A->data;
2229: PetscBLASIntCast(x->nz,&bnz);
2230: y = (Mat_SeqAIJ*)yy->A->data;
2231: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2232: x = (Mat_SeqAIJ*)xx->B->data;
2233: y = (Mat_SeqAIJ*)yy->B->data;
2234: PetscBLASIntCast(x->nz,&bnz);
2235: PetscStackCallBLAS("BLASaxpy",BLASaxpy_(&bnz,&alpha,x->a,&one,y->a,&one));
2236: PetscObjectStateIncrease((PetscObject)Y);
2237: } else if (str == SUBSET_NONZERO_PATTERN) { /* nonzeros of X is a subset of Y's */
2238: MatAXPY_Basic(Y,a,X,str);
2239: } else {
2240: Mat B;
2241: PetscInt *nnz_d,*nnz_o;
2242: PetscMalloc1(yy->A->rmap->N,&nnz_d);
2243: PetscMalloc1(yy->B->rmap->N,&nnz_o);
2244: MatCreate(PetscObjectComm((PetscObject)Y),&B);
2245: PetscObjectSetName((PetscObject)B,((PetscObject)Y)->name);
2246: MatSetSizes(B,Y->rmap->n,Y->cmap->n,Y->rmap->N,Y->cmap->N);
2247: MatSetBlockSizesFromMats(B,Y,Y);
2248: MatSetType(B,MATMPIAIJ);
2249: MatAXPYGetPreallocation_SeqAIJ(yy->A,xx->A,nnz_d);
2250: MatAXPYGetPreallocation_MPIAIJ(yy->B,yy->garray,xx->B,xx->garray,nnz_o);
2251: MatMPIAIJSetPreallocation(B,0,nnz_d,0,nnz_o);
2252: MatAXPY_BasicWithPreallocation(B,Y,a,X,str);
2253: MatHeaderReplace(Y,&B);
2254: PetscFree(nnz_d);
2255: PetscFree(nnz_o);
2256: }
2257: return(0);
2258: }
2260: extern PetscErrorCode MatConjugate_SeqAIJ(Mat);
2262: PetscErrorCode MatConjugate_MPIAIJ(Mat mat)
2263: {
2264: #if defined(PETSC_USE_COMPLEX)
2266: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2269: MatConjugate_SeqAIJ(aij->A);
2270: MatConjugate_SeqAIJ(aij->B);
2271: #else
2273: #endif
2274: return(0);
2275: }
2277: PetscErrorCode MatRealPart_MPIAIJ(Mat A)
2278: {
2279: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2283: MatRealPart(a->A);
2284: MatRealPart(a->B);
2285: return(0);
2286: }
2288: PetscErrorCode MatImaginaryPart_MPIAIJ(Mat A)
2289: {
2290: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2294: MatImaginaryPart(a->A);
2295: MatImaginaryPart(a->B);
2296: return(0);
2297: }
2299: PetscErrorCode MatGetRowMaxAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2300: {
2301: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2303: PetscInt i,*idxb = 0;
2304: PetscScalar *va,*vb;
2305: Vec vtmp;
2308: MatGetRowMaxAbs(a->A,v,idx);
2309: VecGetArray(v,&va);
2310: if (idx) {
2311: for (i=0; i<A->rmap->n; i++) {
2312: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2313: }
2314: }
2316: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2317: if (idx) {
2318: PetscMalloc1(A->rmap->n,&idxb);
2319: }
2320: MatGetRowMaxAbs(a->B,vtmp,idxb);
2321: VecGetArray(vtmp,&vb);
2323: for (i=0; i<A->rmap->n; i++) {
2324: if (PetscAbsScalar(va[i]) < PetscAbsScalar(vb[i])) {
2325: va[i] = vb[i];
2326: if (idx) idx[i] = a->garray[idxb[i]];
2327: }
2328: }
2330: VecRestoreArray(v,&va);
2331: VecRestoreArray(vtmp,&vb);
2332: PetscFree(idxb);
2333: VecDestroy(&vtmp);
2334: return(0);
2335: }
2337: PetscErrorCode MatGetRowMinAbs_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2338: {
2339: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2341: PetscInt i,*idxb = 0;
2342: PetscScalar *va,*vb;
2343: Vec vtmp;
2346: MatGetRowMinAbs(a->A,v,idx);
2347: VecGetArray(v,&va);
2348: if (idx) {
2349: for (i=0; i<A->cmap->n; i++) {
2350: if (PetscAbsScalar(va[i])) idx[i] += A->cmap->rstart;
2351: }
2352: }
2354: VecCreateSeq(PETSC_COMM_SELF,A->rmap->n,&vtmp);
2355: if (idx) {
2356: PetscMalloc1(A->rmap->n,&idxb);
2357: }
2358: MatGetRowMinAbs(a->B,vtmp,idxb);
2359: VecGetArray(vtmp,&vb);
2361: for (i=0; i<A->rmap->n; i++) {
2362: if (PetscAbsScalar(va[i]) > PetscAbsScalar(vb[i])) {
2363: va[i] = vb[i];
2364: if (idx) idx[i] = a->garray[idxb[i]];
2365: }
2366: }
2368: VecRestoreArray(v,&va);
2369: VecRestoreArray(vtmp,&vb);
2370: PetscFree(idxb);
2371: VecDestroy(&vtmp);
2372: return(0);
2373: }
2375: PetscErrorCode MatGetRowMin_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2376: {
2377: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2378: PetscInt n = A->rmap->n;
2379: PetscInt cstart = A->cmap->rstart;
2380: PetscInt *cmap = mat->garray;
2381: PetscInt *diagIdx, *offdiagIdx;
2382: Vec diagV, offdiagV;
2383: PetscScalar *a, *diagA, *offdiagA;
2384: PetscInt r;
2388: PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2389: VecCreateSeq(PetscObjectComm((PetscObject)A), n, &diagV);
2390: VecCreateSeq(PetscObjectComm((PetscObject)A), n, &offdiagV);
2391: MatGetRowMin(mat->A, diagV, diagIdx);
2392: MatGetRowMin(mat->B, offdiagV, offdiagIdx);
2393: VecGetArray(v, &a);
2394: VecGetArray(diagV, &diagA);
2395: VecGetArray(offdiagV, &offdiagA);
2396: for (r = 0; r < n; ++r) {
2397: if (PetscAbsScalar(diagA[r]) <= PetscAbsScalar(offdiagA[r])) {
2398: a[r] = diagA[r];
2399: idx[r] = cstart + diagIdx[r];
2400: } else {
2401: a[r] = offdiagA[r];
2402: idx[r] = cmap[offdiagIdx[r]];
2403: }
2404: }
2405: VecRestoreArray(v, &a);
2406: VecRestoreArray(diagV, &diagA);
2407: VecRestoreArray(offdiagV, &offdiagA);
2408: VecDestroy(&diagV);
2409: VecDestroy(&offdiagV);
2410: PetscFree2(diagIdx, offdiagIdx);
2411: return(0);
2412: }
2414: PetscErrorCode MatGetRowMax_MPIAIJ(Mat A, Vec v, PetscInt idx[])
2415: {
2416: Mat_MPIAIJ *mat = (Mat_MPIAIJ*) A->data;
2417: PetscInt n = A->rmap->n;
2418: PetscInt cstart = A->cmap->rstart;
2419: PetscInt *cmap = mat->garray;
2420: PetscInt *diagIdx, *offdiagIdx;
2421: Vec diagV, offdiagV;
2422: PetscScalar *a, *diagA, *offdiagA;
2423: PetscInt r;
2427: PetscMalloc2(n,&diagIdx,n,&offdiagIdx);
2428: VecCreateSeq(PETSC_COMM_SELF, n, &diagV);
2429: VecCreateSeq(PETSC_COMM_SELF, n, &offdiagV);
2430: MatGetRowMax(mat->A, diagV, diagIdx);
2431: MatGetRowMax(mat->B, offdiagV, offdiagIdx);
2432: VecGetArray(v, &a);
2433: VecGetArray(diagV, &diagA);
2434: VecGetArray(offdiagV, &offdiagA);
2435: for (r = 0; r < n; ++r) {
2436: if (PetscAbsScalar(diagA[r]) >= PetscAbsScalar(offdiagA[r])) {
2437: a[r] = diagA[r];
2438: idx[r] = cstart + diagIdx[r];
2439: } else {
2440: a[r] = offdiagA[r];
2441: idx[r] = cmap[offdiagIdx[r]];
2442: }
2443: }
2444: VecRestoreArray(v, &a);
2445: VecRestoreArray(diagV, &diagA);
2446: VecRestoreArray(offdiagV, &offdiagA);
2447: VecDestroy(&diagV);
2448: VecDestroy(&offdiagV);
2449: PetscFree2(diagIdx, offdiagIdx);
2450: return(0);
2451: }
2453: PetscErrorCode MatGetSeqNonzeroStructure_MPIAIJ(Mat mat,Mat *newmat)
2454: {
2456: Mat *dummy;
2459: MatCreateSubMatrix_MPIAIJ_All(mat,MAT_DO_NOT_GET_VALUES,MAT_INITIAL_MATRIX,&dummy);
2460: *newmat = *dummy;
2461: PetscFree(dummy);
2462: return(0);
2463: }
2465: PetscErrorCode MatInvertBlockDiagonal_MPIAIJ(Mat A,const PetscScalar **values)
2466: {
2467: Mat_MPIAIJ *a = (Mat_MPIAIJ*) A->data;
2471: MatInvertBlockDiagonal(a->A,values);
2472: A->factorerrortype = a->A->factorerrortype;
2473: return(0);
2474: }
2476: static PetscErrorCode MatSetRandom_MPIAIJ(Mat x,PetscRandom rctx)
2477: {
2479: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)x->data;
2482: if (!x->assembled && !x->preallocated) SETERRQ(PetscObjectComm((PetscObject)x), PETSC_ERR_ARG_WRONGSTATE, "MatSetRandom on an unassembled and unpreallocated MATMPIAIJ is not allowed");
2483: MatSetRandom(aij->A,rctx);
2484: if (x->assembled) {
2485: MatSetRandom(aij->B,rctx);
2486: } else {
2487: MatSetRandomSkipColumnRange_SeqAIJ_Private(aij->B,x->cmap->rstart,x->cmap->rend,rctx);
2488: }
2489: MatAssemblyBegin(x,MAT_FINAL_ASSEMBLY);
2490: MatAssemblyEnd(x,MAT_FINAL_ASSEMBLY);
2491: return(0);
2492: }
2494: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ(Mat A,PetscBool sc)
2495: {
2497: if (sc) A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ_Scalable;
2498: else A->ops->increaseoverlap = MatIncreaseOverlap_MPIAIJ;
2499: return(0);
2500: }
2502: /*@
2503: MatMPIAIJSetUseScalableIncreaseOverlap - Determine if the matrix uses a scalable algorithm to compute the overlap
2505: Collective on Mat
2507: Input Parameters:
2508: + A - the matrix
2509: - sc - PETSC_TRUE indicates use the scalable algorithm (default is not to use the scalable algorithm)
2511: Level: advanced
2513: @*/
2514: PetscErrorCode MatMPIAIJSetUseScalableIncreaseOverlap(Mat A,PetscBool sc)
2515: {
2516: PetscErrorCode ierr;
2519: PetscTryMethod(A,"MatMPIAIJSetUseScalableIncreaseOverlap_C",(Mat,PetscBool),(A,sc));
2520: return(0);
2521: }
2523: PetscErrorCode MatSetFromOptions_MPIAIJ(PetscOptionItems *PetscOptionsObject,Mat A)
2524: {
2525: PetscErrorCode ierr;
2526: PetscBool sc = PETSC_FALSE,flg;
2529: PetscOptionsHead(PetscOptionsObject,"MPIAIJ options");
2530: if (A->ops->increaseoverlap == MatIncreaseOverlap_MPIAIJ_Scalable) sc = PETSC_TRUE;
2531: PetscOptionsBool("-mat_increase_overlap_scalable","Use a scalable algorithm to compute the overlap","MatIncreaseOverlap",sc,&sc,&flg);
2532: if (flg) {
2533: MatMPIAIJSetUseScalableIncreaseOverlap(A,sc);
2534: }
2535: PetscOptionsTail();
2536: return(0);
2537: }
2539: PetscErrorCode MatShift_MPIAIJ(Mat Y,PetscScalar a)
2540: {
2542: Mat_MPIAIJ *maij = (Mat_MPIAIJ*)Y->data;
2543: Mat_SeqAIJ *aij = (Mat_SeqAIJ*)maij->A->data;
2546: if (!Y->preallocated) {
2547: MatMPIAIJSetPreallocation(Y,1,NULL,0,NULL);
2548: } else if (!aij->nz) {
2549: PetscInt nonew = aij->nonew;
2550: MatSeqAIJSetPreallocation(maij->A,1,NULL);
2551: aij->nonew = nonew;
2552: }
2553: MatShift_Basic(Y,a);
2554: return(0);
2555: }
2557: PetscErrorCode MatMissingDiagonal_MPIAIJ(Mat A,PetscBool *missing,PetscInt *d)
2558: {
2559: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2563: if (A->rmap->n != A->cmap->n) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"Only works for square matrices");
2564: MatMissingDiagonal(a->A,missing,d);
2565: if (d) {
2566: PetscInt rstart;
2567: MatGetOwnershipRange(A,&rstart,NULL);
2568: *d += rstart;
2570: }
2571: return(0);
2572: }
2574: PetscErrorCode MatInvertVariableBlockDiagonal_MPIAIJ(Mat A,PetscInt nblocks,const PetscInt *bsizes,PetscScalar *diag)
2575: {
2576: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
2580: MatInvertVariableBlockDiagonal(a->A,nblocks,bsizes,diag);
2581: return(0);
2582: }
2584: /* -------------------------------------------------------------------*/
2585: static struct _MatOps MatOps_Values = {MatSetValues_MPIAIJ,
2586: MatGetRow_MPIAIJ,
2587: MatRestoreRow_MPIAIJ,
2588: MatMult_MPIAIJ,
2589: /* 4*/ MatMultAdd_MPIAIJ,
2590: MatMultTranspose_MPIAIJ,
2591: MatMultTransposeAdd_MPIAIJ,
2592: 0,
2593: 0,
2594: 0,
2595: /*10*/ 0,
2596: 0,
2597: 0,
2598: MatSOR_MPIAIJ,
2599: MatTranspose_MPIAIJ,
2600: /*15*/ MatGetInfo_MPIAIJ,
2601: MatEqual_MPIAIJ,
2602: MatGetDiagonal_MPIAIJ,
2603: MatDiagonalScale_MPIAIJ,
2604: MatNorm_MPIAIJ,
2605: /*20*/ MatAssemblyBegin_MPIAIJ,
2606: MatAssemblyEnd_MPIAIJ,
2607: MatSetOption_MPIAIJ,
2608: MatZeroEntries_MPIAIJ,
2609: /*24*/ MatZeroRows_MPIAIJ,
2610: 0,
2611: 0,
2612: 0,
2613: 0,
2614: /*29*/ MatSetUp_MPIAIJ,
2615: 0,
2616: 0,
2617: MatGetDiagonalBlock_MPIAIJ,
2618: 0,
2619: /*34*/ MatDuplicate_MPIAIJ,
2620: 0,
2621: 0,
2622: 0,
2623: 0,
2624: /*39*/ MatAXPY_MPIAIJ,
2625: MatCreateSubMatrices_MPIAIJ,
2626: MatIncreaseOverlap_MPIAIJ,
2627: MatGetValues_MPIAIJ,
2628: MatCopy_MPIAIJ,
2629: /*44*/ MatGetRowMax_MPIAIJ,
2630: MatScale_MPIAIJ,
2631: MatShift_MPIAIJ,
2632: MatDiagonalSet_MPIAIJ,
2633: MatZeroRowsColumns_MPIAIJ,
2634: /*49*/ MatSetRandom_MPIAIJ,
2635: 0,
2636: 0,
2637: 0,
2638: 0,
2639: /*54*/ MatFDColoringCreate_MPIXAIJ,
2640: 0,
2641: MatSetUnfactored_MPIAIJ,
2642: MatPermute_MPIAIJ,
2643: 0,
2644: /*59*/ MatCreateSubMatrix_MPIAIJ,
2645: MatDestroy_MPIAIJ,
2646: MatView_MPIAIJ,
2647: 0,
2648: MatMatMatMult_MPIAIJ_MPIAIJ_MPIAIJ,
2649: /*64*/ MatMatMatMultSymbolic_MPIAIJ_MPIAIJ_MPIAIJ,
2650: MatMatMatMultNumeric_MPIAIJ_MPIAIJ_MPIAIJ,
2651: 0,
2652: 0,
2653: 0,
2654: /*69*/ MatGetRowMaxAbs_MPIAIJ,
2655: MatGetRowMinAbs_MPIAIJ,
2656: 0,
2657: 0,
2658: 0,
2659: 0,
2660: /*75*/ MatFDColoringApply_AIJ,
2661: MatSetFromOptions_MPIAIJ,
2662: 0,
2663: 0,
2664: MatFindZeroDiagonals_MPIAIJ,
2665: /*80*/ 0,
2666: 0,
2667: 0,
2668: /*83*/ MatLoad_MPIAIJ,
2669: MatIsSymmetric_MPIAIJ,
2670: 0,
2671: 0,
2672: 0,
2673: 0,
2674: /*89*/ MatMatMult_MPIAIJ_MPIAIJ,
2675: MatMatMultSymbolic_MPIAIJ_MPIAIJ,
2676: MatMatMultNumeric_MPIAIJ_MPIAIJ,
2677: MatPtAP_MPIAIJ_MPIAIJ,
2678: MatPtAPSymbolic_MPIAIJ_MPIAIJ,
2679: /*94*/ MatPtAPNumeric_MPIAIJ_MPIAIJ,
2680: 0,
2681: 0,
2682: 0,
2683: 0,
2684: /*99*/ 0,
2685: 0,
2686: 0,
2687: MatConjugate_MPIAIJ,
2688: 0,
2689: /*104*/MatSetValuesRow_MPIAIJ,
2690: MatRealPart_MPIAIJ,
2691: MatImaginaryPart_MPIAIJ,
2692: 0,
2693: 0,
2694: /*109*/0,
2695: 0,
2696: MatGetRowMin_MPIAIJ,
2697: 0,
2698: MatMissingDiagonal_MPIAIJ,
2699: /*114*/MatGetSeqNonzeroStructure_MPIAIJ,
2700: 0,
2701: MatGetGhosts_MPIAIJ,
2702: 0,
2703: 0,
2704: /*119*/0,
2705: 0,
2706: 0,
2707: 0,
2708: MatGetMultiProcBlock_MPIAIJ,
2709: /*124*/MatFindNonzeroRows_MPIAIJ,
2710: MatGetColumnNorms_MPIAIJ,
2711: MatInvertBlockDiagonal_MPIAIJ,
2712: MatInvertVariableBlockDiagonal_MPIAIJ,
2713: MatCreateSubMatricesMPI_MPIAIJ,
2714: /*129*/0,
2715: MatTransposeMatMult_MPIAIJ_MPIAIJ,
2716: MatTransposeMatMultSymbolic_MPIAIJ_MPIAIJ,
2717: MatTransposeMatMultNumeric_MPIAIJ_MPIAIJ,
2718: 0,
2719: /*134*/0,
2720: 0,
2721: MatRARt_MPIAIJ_MPIAIJ,
2722: 0,
2723: 0,
2724: /*139*/MatSetBlockSizes_MPIAIJ,
2725: 0,
2726: 0,
2727: MatFDColoringSetUp_MPIXAIJ,
2728: MatFindOffBlockDiagonalEntries_MPIAIJ,
2729: /*144*/MatCreateMPIMatConcatenateSeqMat_MPIAIJ
2730: };
2732: /* ----------------------------------------------------------------------------------------*/
2734: PetscErrorCode MatStoreValues_MPIAIJ(Mat mat)
2735: {
2736: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2740: MatStoreValues(aij->A);
2741: MatStoreValues(aij->B);
2742: return(0);
2743: }
2745: PetscErrorCode MatRetrieveValues_MPIAIJ(Mat mat)
2746: {
2747: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
2751: MatRetrieveValues(aij->A);
2752: MatRetrieveValues(aij->B);
2753: return(0);
2754: }
2756: PetscErrorCode MatMPIAIJSetPreallocation_MPIAIJ(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
2757: {
2758: Mat_MPIAIJ *b;
2760: PetscMPIInt size;
2763: PetscLayoutSetUp(B->rmap);
2764: PetscLayoutSetUp(B->cmap);
2765: b = (Mat_MPIAIJ*)B->data;
2767: #if defined(PETSC_USE_CTABLE)
2768: PetscTableDestroy(&b->colmap);
2769: #else
2770: PetscFree(b->colmap);
2771: #endif
2772: PetscFree(b->garray);
2773: VecDestroy(&b->lvec);
2774: VecScatterDestroy(&b->Mvctx);
2776: /* Because the B will have been resized we simply destroy it and create a new one each time */
2777: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
2778: MatDestroy(&b->B);
2779: MatCreate(PETSC_COMM_SELF,&b->B);
2780: MatSetSizes(b->B,B->rmap->n,size > 1 ? B->cmap->N : 0,B->rmap->n,size > 1 ? B->cmap->N : 0);
2781: MatSetBlockSizesFromMats(b->B,B,B);
2782: MatSetType(b->B,MATSEQAIJ);
2783: PetscLogObjectParent((PetscObject)B,(PetscObject)b->B);
2785: if (!B->preallocated) {
2786: MatCreate(PETSC_COMM_SELF,&b->A);
2787: MatSetSizes(b->A,B->rmap->n,B->cmap->n,B->rmap->n,B->cmap->n);
2788: MatSetBlockSizesFromMats(b->A,B,B);
2789: MatSetType(b->A,MATSEQAIJ);
2790: PetscLogObjectParent((PetscObject)B,(PetscObject)b->A);
2791: }
2793: MatSeqAIJSetPreallocation(b->A,d_nz,d_nnz);
2794: MatSeqAIJSetPreallocation(b->B,o_nz,o_nnz);
2795: B->preallocated = PETSC_TRUE;
2796: B->was_assembled = PETSC_FALSE;
2797: B->assembled = PETSC_FALSE;
2798: return(0);
2799: }
2801: PetscErrorCode MatResetPreallocation_MPIAIJ(Mat B)
2802: {
2803: Mat_MPIAIJ *b;
2808: PetscLayoutSetUp(B->rmap);
2809: PetscLayoutSetUp(B->cmap);
2810: b = (Mat_MPIAIJ*)B->data;
2812: #if defined(PETSC_USE_CTABLE)
2813: PetscTableDestroy(&b->colmap);
2814: #else
2815: PetscFree(b->colmap);
2816: #endif
2817: PetscFree(b->garray);
2818: VecDestroy(&b->lvec);
2819: VecScatterDestroy(&b->Mvctx);
2821: MatResetPreallocation(b->A);
2822: MatResetPreallocation(b->B);
2823: B->preallocated = PETSC_TRUE;
2824: B->was_assembled = PETSC_FALSE;
2825: B->assembled = PETSC_FALSE;
2826: return(0);
2827: }
2829: PetscErrorCode MatDuplicate_MPIAIJ(Mat matin,MatDuplicateOption cpvalues,Mat *newmat)
2830: {
2831: Mat mat;
2832: Mat_MPIAIJ *a,*oldmat = (Mat_MPIAIJ*)matin->data;
2836: *newmat = 0;
2837: MatCreate(PetscObjectComm((PetscObject)matin),&mat);
2838: MatSetSizes(mat,matin->rmap->n,matin->cmap->n,matin->rmap->N,matin->cmap->N);
2839: MatSetBlockSizesFromMats(mat,matin,matin);
2840: MatSetType(mat,((PetscObject)matin)->type_name);
2841: a = (Mat_MPIAIJ*)mat->data;
2843: mat->factortype = matin->factortype;
2844: mat->assembled = PETSC_TRUE;
2845: mat->insertmode = NOT_SET_VALUES;
2846: mat->preallocated = PETSC_TRUE;
2848: a->size = oldmat->size;
2849: a->rank = oldmat->rank;
2850: a->donotstash = oldmat->donotstash;
2851: a->roworiented = oldmat->roworiented;
2852: a->rowindices = 0;
2853: a->rowvalues = 0;
2854: a->getrowactive = PETSC_FALSE;
2856: PetscLayoutReference(matin->rmap,&mat->rmap);
2857: PetscLayoutReference(matin->cmap,&mat->cmap);
2859: if (oldmat->colmap) {
2860: #if defined(PETSC_USE_CTABLE)
2861: PetscTableCreateCopy(oldmat->colmap,&a->colmap);
2862: #else
2863: PetscMalloc1(mat->cmap->N,&a->colmap);
2864: PetscLogObjectMemory((PetscObject)mat,(mat->cmap->N)*sizeof(PetscInt));
2865: PetscMemcpy(a->colmap,oldmat->colmap,(mat->cmap->N)*sizeof(PetscInt));
2866: #endif
2867: } else a->colmap = 0;
2868: if (oldmat->garray) {
2869: PetscInt len;
2870: len = oldmat->B->cmap->n;
2871: PetscMalloc1(len+1,&a->garray);
2872: PetscLogObjectMemory((PetscObject)mat,len*sizeof(PetscInt));
2873: if (len) { PetscMemcpy(a->garray,oldmat->garray,len*sizeof(PetscInt)); }
2874: } else a->garray = 0;
2876: VecDuplicate(oldmat->lvec,&a->lvec);
2877: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->lvec);
2878: VecScatterCopy(oldmat->Mvctx,&a->Mvctx);
2879: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx);
2881: if (oldmat->Mvctx_mpi1) {
2882: VecScatterCopy(oldmat->Mvctx_mpi1,&a->Mvctx_mpi1);
2883: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->Mvctx_mpi1);
2884: }
2886: MatDuplicate(oldmat->A,cpvalues,&a->A);
2887: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->A);
2888: MatDuplicate(oldmat->B,cpvalues,&a->B);
2889: PetscLogObjectParent((PetscObject)mat,(PetscObject)a->B);
2890: PetscFunctionListDuplicate(((PetscObject)matin)->qlist,&((PetscObject)mat)->qlist);
2891: *newmat = mat;
2892: return(0);
2893: }
2895: PetscErrorCode MatLoad_MPIAIJ(Mat newMat, PetscViewer viewer)
2896: {
2897: PetscBool isbinary, ishdf5;
2903: /* force binary viewer to load .info file if it has not yet done so */
2904: PetscViewerSetUp(viewer);
2905: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERBINARY,&isbinary);
2906: PetscObjectTypeCompare((PetscObject)viewer,PETSCVIEWERHDF5, &ishdf5);
2907: if (isbinary) {
2908: MatLoad_MPIAIJ_Binary(newMat,viewer);
2909: } else if (ishdf5) {
2910: #if defined(PETSC_HAVE_HDF5)
2911: MatLoad_AIJ_HDF5(newMat,viewer);
2912: #else
2913: SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"HDF5 not supported in this build.\nPlease reconfigure using --download-hdf5");
2914: #endif
2915: } else {
2916: SETERRQ2(PetscObjectComm((PetscObject)newMat),PETSC_ERR_SUP,"Viewer type %s not yet supported for reading %s matrices",((PetscObject)viewer)->type_name,((PetscObject)newMat)->type_name);
2917: }
2918: return(0);
2919: }
2921: PetscErrorCode MatLoad_MPIAIJ_Binary(Mat newMat, PetscViewer viewer)
2922: {
2923: PetscScalar *vals,*svals;
2924: MPI_Comm comm;
2926: PetscMPIInt rank,size,tag = ((PetscObject)viewer)->tag;
2927: PetscInt i,nz,j,rstart,rend,mmax,maxnz = 0;
2928: PetscInt header[4],*rowlengths = 0,M,N,m,*cols;
2929: PetscInt *ourlens = NULL,*procsnz = NULL,*offlens = NULL,jj,*mycols,*smycols;
2930: PetscInt cend,cstart,n,*rowners;
2931: int fd;
2932: PetscInt bs = newMat->rmap->bs;
2935: PetscObjectGetComm((PetscObject)viewer,&comm);
2936: MPI_Comm_size(comm,&size);
2937: MPI_Comm_rank(comm,&rank);
2938: PetscViewerBinaryGetDescriptor(viewer,&fd);
2939: if (!rank) {
2940: PetscBinaryRead(fd,(char*)header,4,PETSC_INT);
2941: if (header[0] != MAT_FILE_CLASSID) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"not matrix object");
2942: if (header[3] < 0) SETERRQ(PetscObjectComm((PetscObject)newMat),PETSC_ERR_FILE_UNEXPECTED,"Matrix stored in special format on disk,cannot load as MATMPIAIJ");
2943: }
2945: PetscOptionsBegin(comm,NULL,"Options for loading MATMPIAIJ matrix","Mat");
2946: PetscOptionsInt("-matload_block_size","Set the blocksize used to store the matrix","MatLoad",bs,&bs,NULL);
2947: PetscOptionsEnd();
2948: if (bs < 0) bs = 1;
2950: MPI_Bcast(header+1,3,MPIU_INT,0,comm);
2951: M = header[1]; N = header[2];
2953: /* If global sizes are set, check if they are consistent with that given in the file */
2954: if (newMat->rmap->N >= 0 && newMat->rmap->N != M) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of rows:Matrix in file has (%D) and input matrix has (%D)",newMat->rmap->N,M);
2955: if (newMat->cmap->N >=0 && newMat->cmap->N != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED,"Inconsistent # of cols:Matrix in file has (%D) and input matrix has (%D)",newMat->cmap->N,N);
2957: /* determine ownership of all (block) rows */
2958: if (M%bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_FILE_UNEXPECTED, "Inconsistent # of rows (%d) and block size (%d)",M,bs);
2959: if (newMat->rmap->n < 0) m = bs*((M/bs)/size + (((M/bs) % size) > rank)); /* PETSC_DECIDE */
2960: else m = newMat->rmap->n; /* Set by user */
2962: PetscMalloc1(size+1,&rowners);
2963: MPI_Allgather(&m,1,MPIU_INT,rowners+1,1,MPIU_INT,comm);
2965: /* First process needs enough room for process with most rows */
2966: if (!rank) {
2967: mmax = rowners[1];
2968: for (i=2; i<=size; i++) {
2969: mmax = PetscMax(mmax, rowners[i]);
2970: }
2971: } else mmax = -1; /* unused, but compilers complain */
2973: rowners[0] = 0;
2974: for (i=2; i<=size; i++) {
2975: rowners[i] += rowners[i-1];
2976: }
2977: rstart = rowners[rank];
2978: rend = rowners[rank+1];
2980: /* distribute row lengths to all processors */
2981: PetscMalloc2(m,&ourlens,m,&offlens);
2982: if (!rank) {
2983: PetscBinaryRead(fd,ourlens,m,PETSC_INT);
2984: PetscMalloc1(mmax,&rowlengths);
2985: PetscCalloc1(size,&procsnz);
2986: for (j=0; j<m; j++) {
2987: procsnz[0] += ourlens[j];
2988: }
2989: for (i=1; i<size; i++) {
2990: PetscBinaryRead(fd,rowlengths,rowners[i+1]-rowners[i],PETSC_INT);
2991: /* calculate the number of nonzeros on each processor */
2992: for (j=0; j<rowners[i+1]-rowners[i]; j++) {
2993: procsnz[i] += rowlengths[j];
2994: }
2995: MPIULong_Send(rowlengths,rowners[i+1]-rowners[i],MPIU_INT,i,tag,comm);
2996: }
2997: PetscFree(rowlengths);
2998: } else {
2999: MPIULong_Recv(ourlens,m,MPIU_INT,0,tag,comm);
3000: }
3002: if (!rank) {
3003: /* determine max buffer needed and allocate it */
3004: maxnz = 0;
3005: for (i=0; i<size; i++) {
3006: maxnz = PetscMax(maxnz,procsnz[i]);
3007: }
3008: PetscMalloc1(maxnz,&cols);
3010: /* read in my part of the matrix column indices */
3011: nz = procsnz[0];
3012: PetscMalloc1(nz,&mycols);
3013: PetscBinaryRead(fd,mycols,nz,PETSC_INT);
3015: /* read in every one elses and ship off */
3016: for (i=1; i<size; i++) {
3017: nz = procsnz[i];
3018: PetscBinaryRead(fd,cols,nz,PETSC_INT);
3019: MPIULong_Send(cols,nz,MPIU_INT,i,tag,comm);
3020: }
3021: PetscFree(cols);
3022: } else {
3023: /* determine buffer space needed for message */
3024: nz = 0;
3025: for (i=0; i<m; i++) {
3026: nz += ourlens[i];
3027: }
3028: PetscMalloc1(nz,&mycols);
3030: /* receive message of column indices*/
3031: MPIULong_Recv(mycols,nz,MPIU_INT,0,tag,comm);
3032: }
3034: /* determine column ownership if matrix is not square */
3035: if (N != M) {
3036: if (newMat->cmap->n < 0) n = N/size + ((N % size) > rank);
3037: else n = newMat->cmap->n;
3038: MPI_Scan(&n,&cend,1,MPIU_INT,MPI_SUM,comm);
3039: cstart = cend - n;
3040: } else {
3041: cstart = rstart;
3042: cend = rend;
3043: n = cend - cstart;
3044: }
3046: /* loop over local rows, determining number of off diagonal entries */
3047: PetscMemzero(offlens,m*sizeof(PetscInt));
3048: jj = 0;
3049: for (i=0; i<m; i++) {
3050: for (j=0; j<ourlens[i]; j++) {
3051: if (mycols[jj] < cstart || mycols[jj] >= cend) offlens[i]++;
3052: jj++;
3053: }
3054: }
3056: for (i=0; i<m; i++) {
3057: ourlens[i] -= offlens[i];
3058: }
3059: MatSetSizes(newMat,m,n,M,N);
3061: if (bs > 1) {MatSetBlockSize(newMat,bs);}
3063: MatMPIAIJSetPreallocation(newMat,0,ourlens,0,offlens);
3065: for (i=0; i<m; i++) {
3066: ourlens[i] += offlens[i];
3067: }
3069: if (!rank) {
3070: PetscMalloc1(maxnz+1,&vals);
3072: /* read in my part of the matrix numerical values */
3073: nz = procsnz[0];
3074: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3076: /* insert into matrix */
3077: jj = rstart;
3078: smycols = mycols;
3079: svals = vals;
3080: for (i=0; i<m; i++) {
3081: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3082: smycols += ourlens[i];
3083: svals += ourlens[i];
3084: jj++;
3085: }
3087: /* read in other processors and ship out */
3088: for (i=1; i<size; i++) {
3089: nz = procsnz[i];
3090: PetscBinaryRead(fd,vals,nz,PETSC_SCALAR);
3091: MPIULong_Send(vals,nz,MPIU_SCALAR,i,((PetscObject)newMat)->tag,comm);
3092: }
3093: PetscFree(procsnz);
3094: } else {
3095: /* receive numeric values */
3096: PetscMalloc1(nz+1,&vals);
3098: /* receive message of values*/
3099: MPIULong_Recv(vals,nz,MPIU_SCALAR,0,((PetscObject)newMat)->tag,comm);
3101: /* insert into matrix */
3102: jj = rstart;
3103: smycols = mycols;
3104: svals = vals;
3105: for (i=0; i<m; i++) {
3106: MatSetValues_MPIAIJ(newMat,1,&jj,ourlens[i],smycols,svals,INSERT_VALUES);
3107: smycols += ourlens[i];
3108: svals += ourlens[i];
3109: jj++;
3110: }
3111: }
3112: PetscFree2(ourlens,offlens);
3113: PetscFree(vals);
3114: PetscFree(mycols);
3115: PetscFree(rowners);
3116: MatAssemblyBegin(newMat,MAT_FINAL_ASSEMBLY);
3117: MatAssemblyEnd(newMat,MAT_FINAL_ASSEMBLY);
3118: return(0);
3119: }
3121: /* Not scalable because of ISAllGather() unless getting all columns. */
3122: PetscErrorCode ISGetSeqIS_Private(Mat mat,IS iscol,IS *isseq)
3123: {
3125: IS iscol_local;
3126: PetscBool isstride;
3127: PetscMPIInt lisstride=0,gisstride;
3130: /* check if we are grabbing all columns*/
3131: PetscObjectTypeCompare((PetscObject)iscol,ISSTRIDE,&isstride);
3133: if (isstride) {
3134: PetscInt start,len,mstart,mlen;
3135: ISStrideGetInfo(iscol,&start,NULL);
3136: ISGetLocalSize(iscol,&len);
3137: MatGetOwnershipRangeColumn(mat,&mstart,&mlen);
3138: if (mstart == start && mlen-mstart == len) lisstride = 1;
3139: }
3141: MPIU_Allreduce(&lisstride,&gisstride,1,MPI_INT,MPI_MIN,PetscObjectComm((PetscObject)mat));
3142: if (gisstride) {
3143: PetscInt N;
3144: MatGetSize(mat,NULL,&N);
3145: ISCreateStride(PetscObjectComm((PetscObject)mat),N,0,1,&iscol_local);
3146: ISSetIdentity(iscol_local);
3147: PetscInfo(mat,"Optimizing for obtaining all columns of the matrix; skipping ISAllGather()\n");
3148: } else {
3149: PetscInt cbs;
3150: ISGetBlockSize(iscol,&cbs);
3151: ISAllGather(iscol,&iscol_local);
3152: ISSetBlockSize(iscol_local,cbs);
3153: }
3155: *isseq = iscol_local;
3156: return(0);
3157: }
3159: /*
3160: Used by MatCreateSubMatrix_MPIAIJ_SameRowColDist() to avoid ISAllGather() and global size of iscol_local
3161: (see MatCreateSubMatrix_MPIAIJ_nonscalable)
3163: Input Parameters:
3164: mat - matrix
3165: isrow - parallel row index set; its local indices are a subset of local columns of mat,
3166: i.e., mat->rstart <= isrow[i] < mat->rend
3167: iscol - parallel column index set; its local indices are a subset of local columns of mat,
3168: i.e., mat->cstart <= iscol[i] < mat->cend
3169: Output Parameter:
3170: isrow_d,iscol_d - sequential row and column index sets for retrieving mat->A
3171: iscol_o - sequential column index set for retrieving mat->B
3172: garray - column map; garray[i] indicates global location of iscol_o[i] in iscol
3173: */
3174: PetscErrorCode ISGetSeqIS_SameColDist_Private(Mat mat,IS isrow,IS iscol,IS *isrow_d,IS *iscol_d,IS *iscol_o,const PetscInt *garray[])
3175: {
3177: Vec x,cmap;
3178: const PetscInt *is_idx;
3179: PetscScalar *xarray,*cmaparray;
3180: PetscInt ncols,isstart,*idx,m,rstart,*cmap1,count;
3181: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3182: Mat B=a->B;
3183: Vec lvec=a->lvec,lcmap;
3184: PetscInt i,cstart,cend,Bn=B->cmap->N;
3185: MPI_Comm comm;
3186: VecScatter Mvctx=a->Mvctx;
3189: PetscObjectGetComm((PetscObject)mat,&comm);
3190: ISGetLocalSize(iscol,&ncols);
3192: /* (1) iscol is a sub-column vector of mat, pad it with '-1.' to form a full vector x */
3193: MatCreateVecs(mat,&x,NULL);
3194: VecSet(x,-1.0);
3195: VecDuplicate(x,&cmap);
3196: VecSet(cmap,-1.0);
3198: /* Get start indices */
3199: MPI_Scan(&ncols,&isstart,1,MPIU_INT,MPI_SUM,comm);
3200: isstart -= ncols;
3201: MatGetOwnershipRangeColumn(mat,&cstart,&cend);
3203: ISGetIndices(iscol,&is_idx);
3204: VecGetArray(x,&xarray);
3205: VecGetArray(cmap,&cmaparray);
3206: PetscMalloc1(ncols,&idx);
3207: for (i=0; i<ncols; i++) {
3208: xarray[is_idx[i]-cstart] = (PetscScalar)is_idx[i];
3209: cmaparray[is_idx[i]-cstart] = i + isstart; /* global index of iscol[i] */
3210: idx[i] = is_idx[i]-cstart; /* local index of iscol[i] */
3211: }
3212: VecRestoreArray(x,&xarray);
3213: VecRestoreArray(cmap,&cmaparray);
3214: ISRestoreIndices(iscol,&is_idx);
3216: /* Get iscol_d */
3217: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,iscol_d);
3218: ISGetBlockSize(iscol,&i);
3219: ISSetBlockSize(*iscol_d,i);
3221: /* Get isrow_d */
3222: ISGetLocalSize(isrow,&m);
3223: rstart = mat->rmap->rstart;
3224: PetscMalloc1(m,&idx);
3225: ISGetIndices(isrow,&is_idx);
3226: for (i=0; i<m; i++) idx[i] = is_idx[i]-rstart;
3227: ISRestoreIndices(isrow,&is_idx);
3229: ISCreateGeneral(PETSC_COMM_SELF,m,idx,PETSC_OWN_POINTER,isrow_d);
3230: ISGetBlockSize(isrow,&i);
3231: ISSetBlockSize(*isrow_d,i);
3233: /* (2) Scatter x and cmap using aij->Mvctx to get their off-process portions (see MatMult_MPIAIJ) */
3234: VecScatterBegin(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3235: VecScatterEnd(Mvctx,x,lvec,INSERT_VALUES,SCATTER_FORWARD);
3237: VecDuplicate(lvec,&lcmap);
3239: VecScatterBegin(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3240: VecScatterEnd(Mvctx,cmap,lcmap,INSERT_VALUES,SCATTER_FORWARD);
3242: /* (3) create sequential iscol_o (a subset of iscol) and isgarray */
3243: /* off-process column indices */
3244: count = 0;
3245: PetscMalloc1(Bn,&idx);
3246: PetscMalloc1(Bn,&cmap1);
3248: VecGetArray(lvec,&xarray);
3249: VecGetArray(lcmap,&cmaparray);
3250: for (i=0; i<Bn; i++) {
3251: if (PetscRealPart(xarray[i]) > -1.0) {
3252: idx[count] = i; /* local column index in off-diagonal part B */
3253: cmap1[count] = (PetscInt)PetscRealPart(cmaparray[i]); /* column index in submat */
3254: count++;
3255: }
3256: }
3257: VecRestoreArray(lvec,&xarray);
3258: VecRestoreArray(lcmap,&cmaparray);
3260: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_COPY_VALUES,iscol_o);
3261: /* cannot ensure iscol_o has same blocksize as iscol! */
3263: PetscFree(idx);
3264: *garray = cmap1;
3266: VecDestroy(&x);
3267: VecDestroy(&cmap);
3268: VecDestroy(&lcmap);
3269: return(0);
3270: }
3272: /* isrow and iscol have same processor distribution as mat, output *submat is a submatrix of local mat */
3273: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowColDist(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *submat)
3274: {
3276: Mat_MPIAIJ *a = (Mat_MPIAIJ*)mat->data,*asub;
3277: Mat M = NULL;
3278: MPI_Comm comm;
3279: IS iscol_d,isrow_d,iscol_o;
3280: Mat Asub = NULL,Bsub = NULL;
3281: PetscInt n;
3284: PetscObjectGetComm((PetscObject)mat,&comm);
3286: if (call == MAT_REUSE_MATRIX) {
3287: /* Retrieve isrow_d, iscol_d and iscol_o from submat */
3288: PetscObjectQuery((PetscObject)*submat,"isrow_d",(PetscObject*)&isrow_d);
3289: if (!isrow_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"isrow_d passed in was not used before, cannot reuse");
3291: PetscObjectQuery((PetscObject)*submat,"iscol_d",(PetscObject*)&iscol_d);
3292: if (!iscol_d) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_d passed in was not used before, cannot reuse");
3294: PetscObjectQuery((PetscObject)*submat,"iscol_o",(PetscObject*)&iscol_o);
3295: if (!iscol_o) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"iscol_o passed in was not used before, cannot reuse");
3297: /* Update diagonal and off-diagonal portions of submat */
3298: asub = (Mat_MPIAIJ*)(*submat)->data;
3299: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->A);
3300: ISGetLocalSize(iscol_o,&n);
3301: if (n) {
3302: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_REUSE_MATRIX,&asub->B);
3303: }
3304: MatAssemblyBegin(*submat,MAT_FINAL_ASSEMBLY);
3305: MatAssemblyEnd(*submat,MAT_FINAL_ASSEMBLY);
3307: } else { /* call == MAT_INITIAL_MATRIX) */
3308: const PetscInt *garray;
3309: PetscInt BsubN;
3311: /* Create isrow_d, iscol_d, iscol_o and isgarray (replace isgarray with array?) */
3312: ISGetSeqIS_SameColDist_Private(mat,isrow,iscol,&isrow_d,&iscol_d,&iscol_o,&garray);
3314: /* Create local submatrices Asub and Bsub */
3315: MatCreateSubMatrix_SeqAIJ(a->A,isrow_d,iscol_d,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Asub);
3316: MatCreateSubMatrix_SeqAIJ(a->B,isrow_d,iscol_o,PETSC_DECIDE,MAT_INITIAL_MATRIX,&Bsub);
3318: /* Create submatrix M */
3319: MatCreateMPIAIJWithSeqAIJ(comm,Asub,Bsub,garray,&M);
3321: /* If Bsub has empty columns, compress iscol_o such that it will retrieve condensed Bsub from a->B during reuse */
3322: asub = (Mat_MPIAIJ*)M->data;
3324: ISGetLocalSize(iscol_o,&BsubN);
3325: n = asub->B->cmap->N;
3326: if (BsubN > n) {
3327: /* This case can be tested using ~petsc/src/tao/bound/examples/tutorials/runplate2_3 */
3328: const PetscInt *idx;
3329: PetscInt i,j,*idx_new,*subgarray = asub->garray;
3330: PetscInfo2(M,"submatrix Bn %D != BsubN %D, update iscol_o\n",n,BsubN);
3332: PetscMalloc1(n,&idx_new);
3333: j = 0;
3334: ISGetIndices(iscol_o,&idx);
3335: for (i=0; i<n; i++) {
3336: if (j >= BsubN) break;
3337: while (subgarray[i] > garray[j]) j++;
3339: if (subgarray[i] == garray[j]) {
3340: idx_new[i] = idx[j++];
3341: } else SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"subgarray[%D]=%D cannot < garray[%D]=%D",i,subgarray[i],j,garray[j]);
3342: }
3343: ISRestoreIndices(iscol_o,&idx);
3345: ISDestroy(&iscol_o);
3346: ISCreateGeneral(PETSC_COMM_SELF,n,idx_new,PETSC_OWN_POINTER,&iscol_o);
3348: } else if (BsubN < n) {
3349: SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Columns of Bsub cannot be smaller than B's",BsubN,asub->B->cmap->N);
3350: }
3352: PetscFree(garray);
3353: *submat = M;
3355: /* Save isrow_d, iscol_d and iscol_o used in processor for next request */
3356: PetscObjectCompose((PetscObject)M,"isrow_d",(PetscObject)isrow_d);
3357: ISDestroy(&isrow_d);
3359: PetscObjectCompose((PetscObject)M,"iscol_d",(PetscObject)iscol_d);
3360: ISDestroy(&iscol_d);
3362: PetscObjectCompose((PetscObject)M,"iscol_o",(PetscObject)iscol_o);
3363: ISDestroy(&iscol_o);
3364: }
3365: return(0);
3366: }
3368: PetscErrorCode MatCreateSubMatrix_MPIAIJ(Mat mat,IS isrow,IS iscol,MatReuse call,Mat *newmat)
3369: {
3371: IS iscol_local=NULL,isrow_d;
3372: PetscInt csize;
3373: PetscInt n,i,j,start,end;
3374: PetscBool sameRowDist=PETSC_FALSE,sameDist[2],tsameDist[2];
3375: MPI_Comm comm;
3378: /* If isrow has same processor distribution as mat,
3379: call MatCreateSubMatrix_MPIAIJ_SameRowDist() to avoid using a hash table with global size of iscol */
3380: if (call == MAT_REUSE_MATRIX) {
3381: PetscObjectQuery((PetscObject)*newmat,"isrow_d",(PetscObject*)&isrow_d);
3382: if (isrow_d) {
3383: sameRowDist = PETSC_TRUE;
3384: tsameDist[1] = PETSC_TRUE; /* sameColDist */
3385: } else {
3386: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_local);
3387: if (iscol_local) {
3388: sameRowDist = PETSC_TRUE;
3389: tsameDist[1] = PETSC_FALSE; /* !sameColDist */
3390: }
3391: }
3392: } else {
3393: /* Check if isrow has same processor distribution as mat */
3394: sameDist[0] = PETSC_FALSE;
3395: ISGetLocalSize(isrow,&n);
3396: if (!n) {
3397: sameDist[0] = PETSC_TRUE;
3398: } else {
3399: ISGetMinMax(isrow,&i,&j);
3400: MatGetOwnershipRange(mat,&start,&end);
3401: if (i >= start && j < end) {
3402: sameDist[0] = PETSC_TRUE;
3403: }
3404: }
3406: /* Check if iscol has same processor distribution as mat */
3407: sameDist[1] = PETSC_FALSE;
3408: ISGetLocalSize(iscol,&n);
3409: if (!n) {
3410: sameDist[1] = PETSC_TRUE;
3411: } else {
3412: ISGetMinMax(iscol,&i,&j);
3413: MatGetOwnershipRangeColumn(mat,&start,&end);
3414: if (i >= start && j < end) sameDist[1] = PETSC_TRUE;
3415: }
3417: PetscObjectGetComm((PetscObject)mat,&comm);
3418: MPIU_Allreduce(&sameDist,&tsameDist,2,MPIU_BOOL,MPI_LAND,comm);
3419: sameRowDist = tsameDist[0];
3420: }
3422: if (sameRowDist) {
3423: if (tsameDist[1]) { /* sameRowDist & sameColDist */
3424: /* isrow and iscol have same processor distribution as mat */
3425: MatCreateSubMatrix_MPIAIJ_SameRowColDist(mat,isrow,iscol,call,newmat);
3426: return(0);
3427: } else { /* sameRowDist */
3428: /* isrow has same processor distribution as mat */
3429: if (call == MAT_INITIAL_MATRIX) {
3430: PetscBool sorted;
3431: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3432: ISGetLocalSize(iscol_local,&n); /* local size of iscol_local = global columns of newmat */
3433: ISGetSize(iscol,&i);
3434: if (n != i) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"n %d != size of iscol %d",n,i);
3436: ISSorted(iscol_local,&sorted);
3437: if (sorted) {
3438: /* MatCreateSubMatrix_MPIAIJ_SameRowDist() requires iscol_local be sorted; it can have duplicate indices */
3439: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,iscol_local,MAT_INITIAL_MATRIX,newmat);
3440: return(0);
3441: }
3442: } else { /* call == MAT_REUSE_MATRIX */
3443: IS iscol_sub;
3444: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3445: if (iscol_sub) {
3446: MatCreateSubMatrix_MPIAIJ_SameRowDist(mat,isrow,iscol,NULL,call,newmat);
3447: return(0);
3448: }
3449: }
3450: }
3451: }
3453: /* General case: iscol -> iscol_local which has global size of iscol */
3454: if (call == MAT_REUSE_MATRIX) {
3455: PetscObjectQuery((PetscObject)*newmat,"ISAllGather",(PetscObject*)&iscol_local);
3456: if (!iscol_local) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3457: } else {
3458: if (!iscol_local) {
3459: ISGetSeqIS_Private(mat,iscol,&iscol_local);
3460: }
3461: }
3463: ISGetLocalSize(iscol,&csize);
3464: MatCreateSubMatrix_MPIAIJ_nonscalable(mat,isrow,iscol_local,csize,call,newmat);
3466: if (call == MAT_INITIAL_MATRIX) {
3467: PetscObjectCompose((PetscObject)*newmat,"ISAllGather",(PetscObject)iscol_local);
3468: ISDestroy(&iscol_local);
3469: }
3470: return(0);
3471: }
3473: /*@C
3474: MatCreateMPIAIJWithSeqAIJ - creates a MPIAIJ matrix using SeqAIJ matrices that contain the "diagonal"
3475: and "off-diagonal" part of the matrix in CSR format.
3477: Collective on MPI_Comm
3479: Input Parameters:
3480: + comm - MPI communicator
3481: . A - "diagonal" portion of matrix
3482: . B - "off-diagonal" portion of matrix, may have empty columns, will be destroyed by this routine
3483: - garray - global index of B columns
3485: Output Parameter:
3486: . mat - the matrix, with input A as its local diagonal matrix
3487: Level: advanced
3489: Notes:
3490: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix.
3491: A becomes part of output mat, B is destroyed by this routine. The user cannot use A and B anymore.
3493: .seealso: MatCreateMPIAIJWithSplitArrays()
3494: @*/
3495: PetscErrorCode MatCreateMPIAIJWithSeqAIJ(MPI_Comm comm,Mat A,Mat B,const PetscInt garray[],Mat *mat)
3496: {
3498: Mat_MPIAIJ *maij;
3499: Mat_SeqAIJ *b=(Mat_SeqAIJ*)B->data,*bnew;
3500: PetscInt *oi=b->i,*oj=b->j,i,nz,col;
3501: PetscScalar *oa=b->a;
3502: Mat Bnew;
3503: PetscInt m,n,N;
3506: MatCreate(comm,mat);
3507: MatGetSize(A,&m,&n);
3508: if (m != B->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Am %D != Bm %D",m,B->rmap->N);
3509: if (A->rmap->bs != B->rmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A row bs %D != B row bs %D",A->rmap->bs,B->rmap->bs);
3510: /* remove check below; When B is created using iscol_o from ISGetSeqIS_SameColDist_Private(), its bs may not be same as A */
3511: /* if (A->cmap->bs != B->cmap->bs) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"A column bs %D != B column bs %D",A->cmap->bs,B->cmap->bs); */
3513: /* Get global columns of mat */
3514: MPIU_Allreduce(&n,&N,1,MPIU_INT,MPI_SUM,comm);
3516: MatSetSizes(*mat,m,n,PETSC_DECIDE,N);
3517: MatSetType(*mat,MATMPIAIJ);
3518: MatSetBlockSizes(*mat,A->rmap->bs,A->cmap->bs);
3519: maij = (Mat_MPIAIJ*)(*mat)->data;
3521: (*mat)->preallocated = PETSC_TRUE;
3523: PetscLayoutSetUp((*mat)->rmap);
3524: PetscLayoutSetUp((*mat)->cmap);
3526: /* Set A as diagonal portion of *mat */
3527: maij->A = A;
3529: nz = oi[m];
3530: for (i=0; i<nz; i++) {
3531: col = oj[i];
3532: oj[i] = garray[col];
3533: }
3535: /* Set Bnew as off-diagonal portion of *mat */
3536: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,N,oi,oj,oa,&Bnew);
3537: bnew = (Mat_SeqAIJ*)Bnew->data;
3538: bnew->maxnz = b->maxnz; /* allocated nonzeros of B */
3539: maij->B = Bnew;
3541: if (B->rmap->N != Bnew->rmap->N) SETERRQ2(PETSC_COMM_SELF,0,"BN %d != BnewN %d",B->rmap->N,Bnew->rmap->N);
3543: b->singlemalloc = PETSC_FALSE; /* B arrays are shared by Bnew */
3544: b->free_a = PETSC_FALSE;
3545: b->free_ij = PETSC_FALSE;
3546: MatDestroy(&B);
3548: bnew->singlemalloc = PETSC_TRUE; /* arrays will be freed by MatDestroy(&Bnew) */
3549: bnew->free_a = PETSC_TRUE;
3550: bnew->free_ij = PETSC_TRUE;
3552: /* condense columns of maij->B */
3553: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
3554: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
3555: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
3556: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
3557: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3558: return(0);
3559: }
3561: extern PetscErrorCode MatCreateSubMatrices_MPIAIJ_SingleIS_Local(Mat,PetscInt,const IS[],const IS[],MatReuse,PetscBool,Mat*);
3563: PetscErrorCode MatCreateSubMatrix_MPIAIJ_SameRowDist(Mat mat,IS isrow,IS iscol,IS iscol_local,MatReuse call,Mat *newmat)
3564: {
3566: PetscInt i,m,n,rstart,row,rend,nz,j,bs,cbs;
3567: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3568: Mat_MPIAIJ *a=(Mat_MPIAIJ*)mat->data;
3569: Mat M,Msub,B=a->B;
3570: MatScalar *aa;
3571: Mat_SeqAIJ *aij;
3572: PetscInt *garray = a->garray,*colsub,Ncols;
3573: PetscInt count,Bn=B->cmap->N,cstart=mat->cmap->rstart,cend=mat->cmap->rend;
3574: IS iscol_sub,iscmap;
3575: const PetscInt *is_idx,*cmap;
3576: PetscBool allcolumns=PETSC_FALSE;
3577: MPI_Comm comm;
3580: PetscObjectGetComm((PetscObject)mat,&comm);
3582: if (call == MAT_REUSE_MATRIX) {
3583: PetscObjectQuery((PetscObject)*newmat,"SubIScol",(PetscObject*)&iscol_sub);
3584: if (!iscol_sub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"SubIScol passed in was not used before, cannot reuse");
3585: ISGetLocalSize(iscol_sub,&count);
3587: PetscObjectQuery((PetscObject)*newmat,"Subcmap",(PetscObject*)&iscmap);
3588: if (!iscmap) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Subcmap passed in was not used before, cannot reuse");
3590: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Msub);
3591: if (!Msub) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3593: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_REUSE_MATRIX,PETSC_FALSE,&Msub);
3595: } else { /* call == MAT_INITIAL_MATRIX) */
3596: PetscBool flg;
3598: ISGetLocalSize(iscol,&n);
3599: ISGetSize(iscol,&Ncols);
3601: /* (1) iscol -> nonscalable iscol_local */
3602: /* Check for special case: each processor gets entire matrix columns */
3603: ISIdentity(iscol_local,&flg);
3604: if (flg && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3605: if (allcolumns) {
3606: iscol_sub = iscol_local;
3607: PetscObjectReference((PetscObject)iscol_local);
3608: ISCreateStride(PETSC_COMM_SELF,n,0,1,&iscmap);
3610: } else {
3611: /* (2) iscol_local -> iscol_sub and iscmap. Implementation below requires iscol_local be sorted, it can have duplicate indices */
3612: PetscInt *idx,*cmap1,k;
3613: PetscMalloc1(Ncols,&idx);
3614: PetscMalloc1(Ncols,&cmap1);
3615: ISGetIndices(iscol_local,&is_idx);
3616: count = 0;
3617: k = 0;
3618: for (i=0; i<Ncols; i++) {
3619: j = is_idx[i];
3620: if (j >= cstart && j < cend) {
3621: /* diagonal part of mat */
3622: idx[count] = j;
3623: cmap1[count++] = i; /* column index in submat */
3624: } else if (Bn) {
3625: /* off-diagonal part of mat */
3626: if (j == garray[k]) {
3627: idx[count] = j;
3628: cmap1[count++] = i; /* column index in submat */
3629: } else if (j > garray[k]) {
3630: while (j > garray[k] && k < Bn-1) k++;
3631: if (j == garray[k]) {
3632: idx[count] = j;
3633: cmap1[count++] = i; /* column index in submat */
3634: }
3635: }
3636: }
3637: }
3638: ISRestoreIndices(iscol_local,&is_idx);
3640: ISCreateGeneral(PETSC_COMM_SELF,count,idx,PETSC_OWN_POINTER,&iscol_sub);
3641: ISGetBlockSize(iscol,&cbs);
3642: ISSetBlockSize(iscol_sub,cbs);
3644: ISCreateGeneral(PetscObjectComm((PetscObject)iscol_local),count,cmap1,PETSC_OWN_POINTER,&iscmap);
3645: }
3647: /* (3) Create sequential Msub */
3648: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol_sub,MAT_INITIAL_MATRIX,allcolumns,&Msub);
3649: }
3651: ISGetLocalSize(iscol_sub,&count);
3652: aij = (Mat_SeqAIJ*)(Msub)->data;
3653: ii = aij->i;
3654: ISGetIndices(iscmap,&cmap);
3656: /*
3657: m - number of local rows
3658: Ncols - number of columns (same on all processors)
3659: rstart - first row in new global matrix generated
3660: */
3661: MatGetSize(Msub,&m,NULL);
3663: if (call == MAT_INITIAL_MATRIX) {
3664: /* (4) Create parallel newmat */
3665: PetscMPIInt rank,size;
3666: PetscInt csize;
3668: MPI_Comm_size(comm,&size);
3669: MPI_Comm_rank(comm,&rank);
3671: /*
3672: Determine the number of non-zeros in the diagonal and off-diagonal
3673: portions of the matrix in order to do correct preallocation
3674: */
3676: /* first get start and end of "diagonal" columns */
3677: ISGetLocalSize(iscol,&csize);
3678: if (csize == PETSC_DECIDE) {
3679: ISGetSize(isrow,&mglobal);
3680: if (mglobal == Ncols) { /* square matrix */
3681: nlocal = m;
3682: } else {
3683: nlocal = Ncols/size + ((Ncols % size) > rank);
3684: }
3685: } else {
3686: nlocal = csize;
3687: }
3688: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3689: rstart = rend - nlocal;
3690: if (rank == size - 1 && rend != Ncols) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,Ncols);
3692: /* next, compute all the lengths */
3693: jj = aij->j;
3694: PetscMalloc1(2*m+1,&dlens);
3695: olens = dlens + m;
3696: for (i=0; i<m; i++) {
3697: jend = ii[i+1] - ii[i];
3698: olen = 0;
3699: dlen = 0;
3700: for (j=0; j<jend; j++) {
3701: if (cmap[*jj] < rstart || cmap[*jj] >= rend) olen++;
3702: else dlen++;
3703: jj++;
3704: }
3705: olens[i] = olen;
3706: dlens[i] = dlen;
3707: }
3709: ISGetBlockSize(isrow,&bs);
3710: ISGetBlockSize(iscol,&cbs);
3712: MatCreate(comm,&M);
3713: MatSetSizes(M,m,nlocal,PETSC_DECIDE,Ncols);
3714: MatSetBlockSizes(M,bs,cbs);
3715: MatSetType(M,((PetscObject)mat)->type_name);
3716: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3717: PetscFree(dlens);
3719: } else { /* call == MAT_REUSE_MATRIX */
3720: M = *newmat;
3721: MatGetLocalSize(M,&i,NULL);
3722: if (i != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3723: MatZeroEntries(M);
3724: /*
3725: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3726: rather than the slower MatSetValues().
3727: */
3728: M->was_assembled = PETSC_TRUE;
3729: M->assembled = PETSC_FALSE;
3730: }
3732: /* (5) Set values of Msub to *newmat */
3733: PetscMalloc1(count,&colsub);
3734: MatGetOwnershipRange(M,&rstart,NULL);
3736: jj = aij->j;
3737: aa = aij->a;
3738: for (i=0; i<m; i++) {
3739: row = rstart + i;
3740: nz = ii[i+1] - ii[i];
3741: for (j=0; j<nz; j++) colsub[j] = cmap[jj[j]];
3742: MatSetValues_MPIAIJ(M,1,&row,nz,colsub,aa,INSERT_VALUES);
3743: jj += nz; aa += nz;
3744: }
3745: ISRestoreIndices(iscmap,&cmap);
3747: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3748: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3750: PetscFree(colsub);
3752: /* save Msub, iscol_sub and iscmap used in processor for next request */
3753: if (call == MAT_INITIAL_MATRIX) {
3754: *newmat = M;
3755: PetscObjectCompose((PetscObject)(*newmat),"SubMatrix",(PetscObject)Msub);
3756: MatDestroy(&Msub);
3758: PetscObjectCompose((PetscObject)(*newmat),"SubIScol",(PetscObject)iscol_sub);
3759: ISDestroy(&iscol_sub);
3761: PetscObjectCompose((PetscObject)(*newmat),"Subcmap",(PetscObject)iscmap);
3762: ISDestroy(&iscmap);
3764: if (iscol_local) {
3765: PetscObjectCompose((PetscObject)(*newmat),"ISAllGather",(PetscObject)iscol_local);
3766: ISDestroy(&iscol_local);
3767: }
3768: }
3769: return(0);
3770: }
3772: /*
3773: Not great since it makes two copies of the submatrix, first an SeqAIJ
3774: in local and then by concatenating the local matrices the end result.
3775: Writing it directly would be much like MatCreateSubMatrices_MPIAIJ()
3777: Note: This requires a sequential iscol with all indices.
3778: */
3779: PetscErrorCode MatCreateSubMatrix_MPIAIJ_nonscalable(Mat mat,IS isrow,IS iscol,PetscInt csize,MatReuse call,Mat *newmat)
3780: {
3782: PetscMPIInt rank,size;
3783: PetscInt i,m,n,rstart,row,rend,nz,*cwork,j,bs,cbs;
3784: PetscInt *ii,*jj,nlocal,*dlens,*olens,dlen,olen,jend,mglobal;
3785: Mat M,Mreuse;
3786: MatScalar *aa,*vwork;
3787: MPI_Comm comm;
3788: Mat_SeqAIJ *aij;
3789: PetscBool colflag,allcolumns=PETSC_FALSE;
3792: PetscObjectGetComm((PetscObject)mat,&comm);
3793: MPI_Comm_rank(comm,&rank);
3794: MPI_Comm_size(comm,&size);
3796: /* Check for special case: each processor gets entire matrix columns */
3797: ISIdentity(iscol,&colflag);
3798: ISGetLocalSize(iscol,&n);
3799: if (colflag && n == mat->cmap->N) allcolumns = PETSC_TRUE;
3801: if (call == MAT_REUSE_MATRIX) {
3802: PetscObjectQuery((PetscObject)*newmat,"SubMatrix",(PetscObject*)&Mreuse);
3803: if (!Mreuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Submatrix passed in was not used before, cannot reuse");
3804: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_REUSE_MATRIX,allcolumns,&Mreuse);
3805: } else {
3806: MatCreateSubMatrices_MPIAIJ_SingleIS_Local(mat,1,&isrow,&iscol,MAT_INITIAL_MATRIX,allcolumns,&Mreuse);
3807: }
3809: /*
3810: m - number of local rows
3811: n - number of columns (same on all processors)
3812: rstart - first row in new global matrix generated
3813: */
3814: MatGetSize(Mreuse,&m,&n);
3815: MatGetBlockSizes(Mreuse,&bs,&cbs);
3816: if (call == MAT_INITIAL_MATRIX) {
3817: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3818: ii = aij->i;
3819: jj = aij->j;
3821: /*
3822: Determine the number of non-zeros in the diagonal and off-diagonal
3823: portions of the matrix in order to do correct preallocation
3824: */
3826: /* first get start and end of "diagonal" columns */
3827: if (csize == PETSC_DECIDE) {
3828: ISGetSize(isrow,&mglobal);
3829: if (mglobal == n) { /* square matrix */
3830: nlocal = m;
3831: } else {
3832: nlocal = n/size + ((n % size) > rank);
3833: }
3834: } else {
3835: nlocal = csize;
3836: }
3837: MPI_Scan(&nlocal,&rend,1,MPIU_INT,MPI_SUM,comm);
3838: rstart = rend - nlocal;
3839: if (rank == size - 1 && rend != n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Local column sizes %D do not add up to total number of columns %D",rend,n);
3841: /* next, compute all the lengths */
3842: PetscMalloc1(2*m+1,&dlens);
3843: olens = dlens + m;
3844: for (i=0; i<m; i++) {
3845: jend = ii[i+1] - ii[i];
3846: olen = 0;
3847: dlen = 0;
3848: for (j=0; j<jend; j++) {
3849: if (*jj < rstart || *jj >= rend) olen++;
3850: else dlen++;
3851: jj++;
3852: }
3853: olens[i] = olen;
3854: dlens[i] = dlen;
3855: }
3856: MatCreate(comm,&M);
3857: MatSetSizes(M,m,nlocal,PETSC_DECIDE,n);
3858: MatSetBlockSizes(M,bs,cbs);
3859: MatSetType(M,((PetscObject)mat)->type_name);
3860: MatMPIAIJSetPreallocation(M,0,dlens,0,olens);
3861: PetscFree(dlens);
3862: } else {
3863: PetscInt ml,nl;
3865: M = *newmat;
3866: MatGetLocalSize(M,&ml,&nl);
3867: if (ml != m) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Previous matrix must be same size/layout as request");
3868: MatZeroEntries(M);
3869: /*
3870: The next two lines are needed so we may call MatSetValues_MPIAIJ() below directly,
3871: rather than the slower MatSetValues().
3872: */
3873: M->was_assembled = PETSC_TRUE;
3874: M->assembled = PETSC_FALSE;
3875: }
3876: MatGetOwnershipRange(M,&rstart,&rend);
3877: aij = (Mat_SeqAIJ*)(Mreuse)->data;
3878: ii = aij->i;
3879: jj = aij->j;
3880: aa = aij->a;
3881: for (i=0; i<m; i++) {
3882: row = rstart + i;
3883: nz = ii[i+1] - ii[i];
3884: cwork = jj; jj += nz;
3885: vwork = aa; aa += nz;
3886: MatSetValues_MPIAIJ(M,1,&row,nz,cwork,vwork,INSERT_VALUES);
3887: }
3889: MatAssemblyBegin(M,MAT_FINAL_ASSEMBLY);
3890: MatAssemblyEnd(M,MAT_FINAL_ASSEMBLY);
3891: *newmat = M;
3893: /* save submatrix used in processor for next request */
3894: if (call == MAT_INITIAL_MATRIX) {
3895: PetscObjectCompose((PetscObject)M,"SubMatrix",(PetscObject)Mreuse);
3896: MatDestroy(&Mreuse);
3897: }
3898: return(0);
3899: }
3901: PetscErrorCode MatMPIAIJSetPreallocationCSR_MPIAIJ(Mat B,const PetscInt Ii[],const PetscInt J[],const PetscScalar v[])
3902: {
3903: PetscInt m,cstart, cend,j,nnz,i,d;
3904: PetscInt *d_nnz,*o_nnz,nnz_max = 0,rstart,ii;
3905: const PetscInt *JJ;
3906: PetscScalar *values;
3908: PetscBool nooffprocentries;
3911: if (Ii && Ii[0]) SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Ii[0] must be 0 it is %D",Ii[0]);
3913: PetscLayoutSetUp(B->rmap);
3914: PetscLayoutSetUp(B->cmap);
3915: m = B->rmap->n;
3916: cstart = B->cmap->rstart;
3917: cend = B->cmap->rend;
3918: rstart = B->rmap->rstart;
3920: PetscCalloc2(m,&d_nnz,m,&o_nnz);
3922: #if defined(PETSC_USE_DEBUG)
3923: for (i=0; i<m && Ii; i++) {
3924: nnz = Ii[i+1]- Ii[i];
3925: JJ = J + Ii[i];
3926: if (nnz < 0) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Local row %D has a negative %D number of columns",i,nnz);
3927: if (nnz && (JJ[0] < 0)) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D starts with negative column index",i,JJ[0]);
3928: if (nnz && (JJ[nnz-1] >= B->cmap->N)) SETERRQ3(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Row %D ends with too large a column index %D (max allowed %D)",i,JJ[nnz-1],B->cmap->N);
3929: }
3930: #endif
3932: for (i=0; i<m && Ii; i++) {
3933: nnz = Ii[i+1]- Ii[i];
3934: JJ = J + Ii[i];
3935: nnz_max = PetscMax(nnz_max,nnz);
3936: d = 0;
3937: for (j=0; j<nnz; j++) {
3938: if (cstart <= JJ[j] && JJ[j] < cend) d++;
3939: }
3940: d_nnz[i] = d;
3941: o_nnz[i] = nnz - d;
3942: }
3943: MatMPIAIJSetPreallocation(B,0,d_nnz,0,o_nnz);
3944: PetscFree2(d_nnz,o_nnz);
3946: if (v) values = (PetscScalar*)v;
3947: else {
3948: PetscCalloc1(nnz_max+1,&values);
3949: }
3951: for (i=0; i<m && Ii; i++) {
3952: ii = i + rstart;
3953: nnz = Ii[i+1]- Ii[i];
3954: MatSetValues_MPIAIJ(B,1,&ii,nnz,J+Ii[i],values+(v ? Ii[i] : 0),INSERT_VALUES);
3955: }
3956: nooffprocentries = B->nooffprocentries;
3957: B->nooffprocentries = PETSC_TRUE;
3958: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
3959: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
3960: B->nooffprocentries = nooffprocentries;
3962: if (!v) {
3963: PetscFree(values);
3964: }
3965: MatSetOption(B,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
3966: return(0);
3967: }
3969: /*@
3970: MatMPIAIJSetPreallocationCSR - Allocates memory for a sparse parallel matrix in AIJ format
3971: (the default parallel PETSc format).
3973: Collective on MPI_Comm
3975: Input Parameters:
3976: + B - the matrix
3977: . i - the indices into j for the start of each local row (starts with zero)
3978: . j - the column indices for each local row (starts with zero)
3979: - v - optional values in the matrix
3981: Level: developer
3983: Notes:
3984: The i, j, and v arrays ARE copied by this routine into the internal format used by PETSc;
3985: thus you CANNOT change the matrix entries by changing the values of v[] after you have
3986: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
3988: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
3990: The format which is used for the sparse matrix input, is equivalent to a
3991: row-major ordering.. i.e for the following matrix, the input data expected is
3992: as shown
3994: $ 1 0 0
3995: $ 2 0 3 P0
3996: $ -------
3997: $ 4 5 6 P1
3998: $
3999: $ Process0 [P0]: rows_owned=[0,1]
4000: $ i = {0,1,3} [size = nrow+1 = 2+1]
4001: $ j = {0,0,2} [size = 3]
4002: $ v = {1,2,3} [size = 3]
4003: $
4004: $ Process1 [P1]: rows_owned=[2]
4005: $ i = {0,3} [size = nrow+1 = 1+1]
4006: $ j = {0,1,2} [size = 3]
4007: $ v = {4,5,6} [size = 3]
4009: .keywords: matrix, aij, compressed row, sparse, parallel
4011: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatCreateAIJ(), MATMPIAIJ,
4012: MatCreateSeqAIJWithArrays(), MatCreateMPIAIJWithSplitArrays()
4013: @*/
4014: PetscErrorCode MatMPIAIJSetPreallocationCSR(Mat B,const PetscInt i[],const PetscInt j[], const PetscScalar v[])
4015: {
4019: PetscTryMethod(B,"MatMPIAIJSetPreallocationCSR_C",(Mat,const PetscInt[],const PetscInt[],const PetscScalar[]),(B,i,j,v));
4020: return(0);
4021: }
4023: /*@C
4024: MatMPIAIJSetPreallocation - Preallocates memory for a sparse parallel matrix in AIJ format
4025: (the default parallel PETSc format). For good matrix assembly performance
4026: the user should preallocate the matrix storage by setting the parameters
4027: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4028: performance can be increased by more than a factor of 50.
4030: Collective on MPI_Comm
4032: Input Parameters:
4033: + B - the matrix
4034: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4035: (same value is used for all local rows)
4036: . d_nnz - array containing the number of nonzeros in the various rows of the
4037: DIAGONAL portion of the local submatrix (possibly different for each row)
4038: or NULL (PETSC_NULL_INTEGER in Fortran), if d_nz is used to specify the nonzero structure.
4039: The size of this array is equal to the number of local rows, i.e 'm'.
4040: For matrices that will be factored, you must leave room for (and set)
4041: the diagonal entry even if it is zero.
4042: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4043: submatrix (same value is used for all local rows).
4044: - o_nnz - array containing the number of nonzeros in the various rows of the
4045: OFF-DIAGONAL portion of the local submatrix (possibly different for
4046: each row) or NULL (PETSC_NULL_INTEGER in Fortran), if o_nz is used to specify the nonzero
4047: structure. The size of this array is equal to the number
4048: of local rows, i.e 'm'.
4050: If the *_nnz parameter is given then the *_nz parameter is ignored
4052: The AIJ format (also called the Yale sparse matrix format or
4053: compressed row storage (CSR)), is fully compatible with standard Fortran 77
4054: storage. The stored row and column indices begin with zero.
4055: See Users-Manual: ch_mat for details.
4057: The parallel matrix is partitioned such that the first m0 rows belong to
4058: process 0, the next m1 rows belong to process 1, the next m2 rows belong
4059: to process 2 etc.. where m0,m1,m2... are the input parameter 'm'.
4061: The DIAGONAL portion of the local submatrix of a processor can be defined
4062: as the submatrix which is obtained by extraction the part corresponding to
4063: the rows r1-r2 and columns c1-c2 of the global matrix, where r1 is the
4064: first row that belongs to the processor, r2 is the last row belonging to
4065: the this processor, and c1-c2 is range of indices of the local part of a
4066: vector suitable for applying the matrix to. This is an mxn matrix. In the
4067: common case of a square matrix, the row and column ranges are the same and
4068: the DIAGONAL part is also square. The remaining portion of the local
4069: submatrix (mxN) constitute the OFF-DIAGONAL portion.
4071: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4073: You can call MatGetInfo() to get information on how effective the preallocation was;
4074: for example the fields mallocs,nz_allocated,nz_used,nz_unneeded;
4075: You can also run with the option -info and look for messages with the string
4076: malloc in them to see if additional memory allocation was needed.
4078: Example usage:
4080: Consider the following 8x8 matrix with 34 non-zero values, that is
4081: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4082: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4083: as follows:
4085: .vb
4086: 1 2 0 | 0 3 0 | 0 4
4087: Proc0 0 5 6 | 7 0 0 | 8 0
4088: 9 0 10 | 11 0 0 | 12 0
4089: -------------------------------------
4090: 13 0 14 | 15 16 17 | 0 0
4091: Proc1 0 18 0 | 19 20 21 | 0 0
4092: 0 0 0 | 22 23 0 | 24 0
4093: -------------------------------------
4094: Proc2 25 26 27 | 0 0 28 | 29 0
4095: 30 0 0 | 31 32 33 | 0 34
4096: .ve
4098: This can be represented as a collection of submatrices as:
4100: .vb
4101: A B C
4102: D E F
4103: G H I
4104: .ve
4106: Where the submatrices A,B,C are owned by proc0, D,E,F are
4107: owned by proc1, G,H,I are owned by proc2.
4109: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4110: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4111: The 'M','N' parameters are 8,8, and have the same values on all procs.
4113: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4114: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4115: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4116: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4117: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4118: matrix, ans [DF] as another SeqAIJ matrix.
4120: When d_nz, o_nz parameters are specified, d_nz storage elements are
4121: allocated for every row of the local diagonal submatrix, and o_nz
4122: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4123: One way to choose d_nz and o_nz is to use the max nonzerors per local
4124: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4125: In this case, the values of d_nz,o_nz are:
4126: .vb
4127: proc0 : dnz = 2, o_nz = 2
4128: proc1 : dnz = 3, o_nz = 2
4129: proc2 : dnz = 1, o_nz = 4
4130: .ve
4131: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4132: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4133: for proc3. i.e we are using 12+15+10=37 storage locations to store
4134: 34 values.
4136: When d_nnz, o_nnz parameters are specified, the storage is specified
4137: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4138: In the above case the values for d_nnz,o_nnz are:
4139: .vb
4140: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4141: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4142: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4143: .ve
4144: Here the space allocated is sum of all the above values i.e 34, and
4145: hence pre-allocation is perfect.
4147: Level: intermediate
4149: .keywords: matrix, aij, compressed row, sparse, parallel
4151: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatCreateAIJ(), MatMPIAIJSetPreallocationCSR(),
4152: MATMPIAIJ, MatGetInfo(), PetscSplitOwnership()
4153: @*/
4154: PetscErrorCode MatMPIAIJSetPreallocation(Mat B,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[])
4155: {
4161: PetscTryMethod(B,"MatMPIAIJSetPreallocation_C",(Mat,PetscInt,const PetscInt[],PetscInt,const PetscInt[]),(B,d_nz,d_nnz,o_nz,o_nnz));
4162: return(0);
4163: }
4165: /*@
4166: MatCreateMPIAIJWithArrays - creates a MPI AIJ matrix using arrays that contain in standard
4167: CSR format the local rows.
4169: Collective on MPI_Comm
4171: Input Parameters:
4172: + comm - MPI communicator
4173: . m - number of local rows (Cannot be PETSC_DECIDE)
4174: . n - This value should be the same as the local size used in creating the
4175: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4176: calculated if N is given) For square matrices n is almost always m.
4177: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4178: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4179: . i - row indices; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
4180: . j - column indices
4181: - a - matrix values
4183: Output Parameter:
4184: . mat - the matrix
4186: Level: intermediate
4188: Notes:
4189: The i, j, and a arrays ARE copied by this routine into the internal format used by PETSc;
4190: thus you CANNOT change the matrix entries by changing the values of a[] after you have
4191: called this routine. Use MatCreateMPIAIJWithSplitArrays() to avoid needing to copy the arrays.
4193: The i and j indices are 0 based, and i indices are indices corresponding to the local j array.
4195: The format which is used for the sparse matrix input, is equivalent to a
4196: row-major ordering.. i.e for the following matrix, the input data expected is
4197: as shown
4199: $ 1 0 0
4200: $ 2 0 3 P0
4201: $ -------
4202: $ 4 5 6 P1
4203: $
4204: $ Process0 [P0]: rows_owned=[0,1]
4205: $ i = {0,1,3} [size = nrow+1 = 2+1]
4206: $ j = {0,0,2} [size = 3]
4207: $ v = {1,2,3} [size = 3]
4208: $
4209: $ Process1 [P1]: rows_owned=[2]
4210: $ i = {0,3} [size = nrow+1 = 1+1]
4211: $ j = {0,1,2} [size = 3]
4212: $ v = {4,5,6} [size = 3]
4214: .keywords: matrix, aij, compressed row, sparse, parallel
4216: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4217: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithSplitArrays()
4218: @*/
4219: PetscErrorCode MatCreateMPIAIJWithArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,const PetscInt i[],const PetscInt j[],const PetscScalar a[],Mat *mat)
4220: {
4224: if (i && i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
4225: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
4226: MatCreate(comm,mat);
4227: MatSetSizes(*mat,m,n,M,N);
4228: /* MatSetBlockSizes(M,bs,cbs); */
4229: MatSetType(*mat,MATMPIAIJ);
4230: MatMPIAIJSetPreallocationCSR(*mat,i,j,a);
4231: return(0);
4232: }
4234: /*@C
4235: MatCreateAIJ - Creates a sparse parallel matrix in AIJ format
4236: (the default parallel PETSc format). For good matrix assembly performance
4237: the user should preallocate the matrix storage by setting the parameters
4238: d_nz (or d_nnz) and o_nz (or o_nnz). By setting these parameters accurately,
4239: performance can be increased by more than a factor of 50.
4241: Collective on MPI_Comm
4243: Input Parameters:
4244: + comm - MPI communicator
4245: . m - number of local rows (or PETSC_DECIDE to have calculated if M is given)
4246: This value should be the same as the local size used in creating the
4247: y vector for the matrix-vector product y = Ax.
4248: . n - This value should be the same as the local size used in creating the
4249: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
4250: calculated if N is given) For square matrices n is almost always m.
4251: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
4252: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
4253: . d_nz - number of nonzeros per row in DIAGONAL portion of local submatrix
4254: (same value is used for all local rows)
4255: . d_nnz - array containing the number of nonzeros in the various rows of the
4256: DIAGONAL portion of the local submatrix (possibly different for each row)
4257: or NULL, if d_nz is used to specify the nonzero structure.
4258: The size of this array is equal to the number of local rows, i.e 'm'.
4259: . o_nz - number of nonzeros per row in the OFF-DIAGONAL portion of local
4260: submatrix (same value is used for all local rows).
4261: - o_nnz - array containing the number of nonzeros in the various rows of the
4262: OFF-DIAGONAL portion of the local submatrix (possibly different for
4263: each row) or NULL, if o_nz is used to specify the nonzero
4264: structure. The size of this array is equal to the number
4265: of local rows, i.e 'm'.
4267: Output Parameter:
4268: . A - the matrix
4270: It is recommended that one use the MatCreate(), MatSetType() and/or MatSetFromOptions(),
4271: MatXXXXSetPreallocation() paradigm instead of this routine directly.
4272: [MatXXXXSetPreallocation() is, for example, MatSeqAIJSetPreallocation]
4274: Notes:
4275: If the *_nnz parameter is given then the *_nz parameter is ignored
4277: m,n,M,N parameters specify the size of the matrix, and its partitioning across
4278: processors, while d_nz,d_nnz,o_nz,o_nnz parameters specify the approximate
4279: storage requirements for this matrix.
4281: If PETSC_DECIDE or PETSC_DETERMINE is used for a particular argument on one
4282: processor than it must be used on all processors that share the object for
4283: that argument.
4285: The user MUST specify either the local or global matrix dimensions
4286: (possibly both).
4288: The parallel matrix is partitioned across processors such that the
4289: first m0 rows belong to process 0, the next m1 rows belong to
4290: process 1, the next m2 rows belong to process 2 etc.. where
4291: m0,m1,m2,.. are the input parameter 'm'. i.e each processor stores
4292: values corresponding to [m x N] submatrix.
4294: The columns are logically partitioned with the n0 columns belonging
4295: to 0th partition, the next n1 columns belonging to the next
4296: partition etc.. where n0,n1,n2... are the input parameter 'n'.
4298: The DIAGONAL portion of the local submatrix on any given processor
4299: is the submatrix corresponding to the rows and columns m,n
4300: corresponding to the given processor. i.e diagonal matrix on
4301: process 0 is [m0 x n0], diagonal matrix on process 1 is [m1 x n1]
4302: etc. The remaining portion of the local submatrix [m x (N-n)]
4303: constitute the OFF-DIAGONAL portion. The example below better
4304: illustrates this concept.
4306: For a square global matrix we define each processor's diagonal portion
4307: to be its local rows and the corresponding columns (a square submatrix);
4308: each processor's off-diagonal portion encompasses the remainder of the
4309: local matrix (a rectangular submatrix).
4311: If o_nnz, d_nnz are specified, then o_nz, and d_nz are ignored.
4313: When calling this routine with a single process communicator, a matrix of
4314: type SEQAIJ is returned. If a matrix of type MPIAIJ is desired for this
4315: type of communicator, use the construction mechanism
4316: .vb
4317: MatCreate(...,&A); MatSetType(A,MATMPIAIJ); MatSetSizes(A, m,n,M,N); MatMPIAIJSetPreallocation(A,...);
4318: .ve
4320: $ MatCreate(...,&A);
4321: $ MatSetType(A,MATMPIAIJ);
4322: $ MatSetSizes(A, m,n,M,N);
4323: $ MatMPIAIJSetPreallocation(A,...);
4325: By default, this format uses inodes (identical nodes) when possible.
4326: We search for consecutive rows with the same nonzero structure, thereby
4327: reusing matrix information to achieve increased efficiency.
4329: Options Database Keys:
4330: + -mat_no_inode - Do not use inodes
4331: - -mat_inode_limit <limit> - Sets inode limit (max limit=5)
4335: Example usage:
4337: Consider the following 8x8 matrix with 34 non-zero values, that is
4338: assembled across 3 processors. Lets assume that proc0 owns 3 rows,
4339: proc1 owns 3 rows, proc2 owns 2 rows. This division can be shown
4340: as follows
4342: .vb
4343: 1 2 0 | 0 3 0 | 0 4
4344: Proc0 0 5 6 | 7 0 0 | 8 0
4345: 9 0 10 | 11 0 0 | 12 0
4346: -------------------------------------
4347: 13 0 14 | 15 16 17 | 0 0
4348: Proc1 0 18 0 | 19 20 21 | 0 0
4349: 0 0 0 | 22 23 0 | 24 0
4350: -------------------------------------
4351: Proc2 25 26 27 | 0 0 28 | 29 0
4352: 30 0 0 | 31 32 33 | 0 34
4353: .ve
4355: This can be represented as a collection of submatrices as
4357: .vb
4358: A B C
4359: D E F
4360: G H I
4361: .ve
4363: Where the submatrices A,B,C are owned by proc0, D,E,F are
4364: owned by proc1, G,H,I are owned by proc2.
4366: The 'm' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4367: The 'n' parameters for proc0,proc1,proc2 are 3,3,2 respectively.
4368: The 'M','N' parameters are 8,8, and have the same values on all procs.
4370: The DIAGONAL submatrices corresponding to proc0,proc1,proc2 are
4371: submatrices [A], [E], [I] respectively. The OFF-DIAGONAL submatrices
4372: corresponding to proc0,proc1,proc2 are [BC], [DF], [GH] respectively.
4373: Internally, each processor stores the DIAGONAL part, and the OFF-DIAGONAL
4374: part as SeqAIJ matrices. for eg: proc1 will store [E] as a SeqAIJ
4375: matrix, ans [DF] as another SeqAIJ matrix.
4377: When d_nz, o_nz parameters are specified, d_nz storage elements are
4378: allocated for every row of the local diagonal submatrix, and o_nz
4379: storage locations are allocated for every row of the OFF-DIAGONAL submat.
4380: One way to choose d_nz and o_nz is to use the max nonzerors per local
4381: rows for each of the local DIAGONAL, and the OFF-DIAGONAL submatrices.
4382: In this case, the values of d_nz,o_nz are
4383: .vb
4384: proc0 : dnz = 2, o_nz = 2
4385: proc1 : dnz = 3, o_nz = 2
4386: proc2 : dnz = 1, o_nz = 4
4387: .ve
4388: We are allocating m*(d_nz+o_nz) storage locations for every proc. This
4389: translates to 3*(2+2)=12 for proc0, 3*(3+2)=15 for proc1, 2*(1+4)=10
4390: for proc3. i.e we are using 12+15+10=37 storage locations to store
4391: 34 values.
4393: When d_nnz, o_nnz parameters are specified, the storage is specified
4394: for every row, coresponding to both DIAGONAL and OFF-DIAGONAL submatrices.
4395: In the above case the values for d_nnz,o_nnz are
4396: .vb
4397: proc0: d_nnz = [2,2,2] and o_nnz = [2,2,2]
4398: proc1: d_nnz = [3,3,2] and o_nnz = [2,1,1]
4399: proc2: d_nnz = [1,1] and o_nnz = [4,4]
4400: .ve
4401: Here the space allocated is sum of all the above values i.e 34, and
4402: hence pre-allocation is perfect.
4404: Level: intermediate
4406: .keywords: matrix, aij, compressed row, sparse, parallel
4408: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
4409: MATMPIAIJ, MatCreateMPIAIJWithArrays()
4410: @*/
4411: PetscErrorCode MatCreateAIJ(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt d_nz,const PetscInt d_nnz[],PetscInt o_nz,const PetscInt o_nnz[],Mat *A)
4412: {
4414: PetscMPIInt size;
4417: MatCreate(comm,A);
4418: MatSetSizes(*A,m,n,M,N);
4419: MPI_Comm_size(comm,&size);
4420: if (size > 1) {
4421: MatSetType(*A,MATMPIAIJ);
4422: MatMPIAIJSetPreallocation(*A,d_nz,d_nnz,o_nz,o_nnz);
4423: } else {
4424: MatSetType(*A,MATSEQAIJ);
4425: MatSeqAIJSetPreallocation(*A,d_nz,d_nnz);
4426: }
4427: return(0);
4428: }
4430: PetscErrorCode MatMPIAIJGetSeqAIJ(Mat A,Mat *Ad,Mat *Ao,const PetscInt *colmap[])
4431: {
4432: Mat_MPIAIJ *a = (Mat_MPIAIJ*)A->data;
4433: PetscBool flg;
4437: PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&flg);
4438: if (!flg) SETERRQ(PetscObjectComm((PetscObject)A),PETSC_ERR_SUP,"This function requires a MATMPIAIJ matrix as input");
4439: if (Ad) *Ad = a->A;
4440: if (Ao) *Ao = a->B;
4441: if (colmap) *colmap = a->garray;
4442: return(0);
4443: }
4445: PetscErrorCode MatCreateMPIMatConcatenateSeqMat_MPIAIJ(MPI_Comm comm,Mat inmat,PetscInt n,MatReuse scall,Mat *outmat)
4446: {
4448: PetscInt m,N,i,rstart,nnz,Ii;
4449: PetscInt *indx;
4450: PetscScalar *values;
4453: MatGetSize(inmat,&m,&N);
4454: if (scall == MAT_INITIAL_MATRIX) { /* symbolic phase */
4455: PetscInt *dnz,*onz,sum,bs,cbs;
4457: if (n == PETSC_DECIDE) {
4458: PetscSplitOwnership(comm,&n,&N);
4459: }
4460: /* Check sum(n) = N */
4461: MPIU_Allreduce(&n,&sum,1,MPIU_INT,MPI_SUM,comm);
4462: if (sum != N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_INCOMP,"Sum of local columns %D != global columns %D",sum,N);
4464: MPI_Scan(&m, &rstart,1,MPIU_INT,MPI_SUM,comm);
4465: rstart -= m;
4467: MatPreallocateInitialize(comm,m,n,dnz,onz);
4468: for (i=0; i<m; i++) {
4469: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4470: MatPreallocateSet(i+rstart,nnz,indx,dnz,onz);
4471: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,NULL);
4472: }
4474: MatCreate(comm,outmat);
4475: MatSetSizes(*outmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4476: MatGetBlockSizes(inmat,&bs,&cbs);
4477: MatSetBlockSizes(*outmat,bs,cbs);
4478: MatSetType(*outmat,MATAIJ);
4479: MatSeqAIJSetPreallocation(*outmat,0,dnz);
4480: MatMPIAIJSetPreallocation(*outmat,0,dnz,0,onz);
4481: MatPreallocateFinalize(dnz,onz);
4482: }
4484: /* numeric phase */
4485: MatGetOwnershipRange(*outmat,&rstart,NULL);
4486: for (i=0; i<m; i++) {
4487: MatGetRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4488: Ii = i + rstart;
4489: MatSetValues(*outmat,1,&Ii,nnz,indx,values,INSERT_VALUES);
4490: MatRestoreRow_SeqAIJ(inmat,i,&nnz,&indx,&values);
4491: }
4492: MatAssemblyBegin(*outmat,MAT_FINAL_ASSEMBLY);
4493: MatAssemblyEnd(*outmat,MAT_FINAL_ASSEMBLY);
4494: return(0);
4495: }
4497: PetscErrorCode MatFileSplit(Mat A,char *outfile)
4498: {
4499: PetscErrorCode ierr;
4500: PetscMPIInt rank;
4501: PetscInt m,N,i,rstart,nnz;
4502: size_t len;
4503: const PetscInt *indx;
4504: PetscViewer out;
4505: char *name;
4506: Mat B;
4507: const PetscScalar *values;
4510: MatGetLocalSize(A,&m,0);
4511: MatGetSize(A,0,&N);
4512: /* Should this be the type of the diagonal block of A? */
4513: MatCreate(PETSC_COMM_SELF,&B);
4514: MatSetSizes(B,m,N,m,N);
4515: MatSetBlockSizesFromMats(B,A,A);
4516: MatSetType(B,MATSEQAIJ);
4517: MatSeqAIJSetPreallocation(B,0,NULL);
4518: MatGetOwnershipRange(A,&rstart,0);
4519: for (i=0; i<m; i++) {
4520: MatGetRow(A,i+rstart,&nnz,&indx,&values);
4521: MatSetValues(B,1,&i,nnz,indx,values,INSERT_VALUES);
4522: MatRestoreRow(A,i+rstart,&nnz,&indx,&values);
4523: }
4524: MatAssemblyBegin(B,MAT_FINAL_ASSEMBLY);
4525: MatAssemblyEnd(B,MAT_FINAL_ASSEMBLY);
4527: MPI_Comm_rank(PetscObjectComm((PetscObject)A),&rank);
4528: PetscStrlen(outfile,&len);
4529: PetscMalloc1(len+5,&name);
4530: sprintf(name,"%s.%d",outfile,rank);
4531: PetscViewerBinaryOpen(PETSC_COMM_SELF,name,FILE_MODE_APPEND,&out);
4532: PetscFree(name);
4533: MatView(B,out);
4534: PetscViewerDestroy(&out);
4535: MatDestroy(&B);
4536: return(0);
4537: }
4539: PetscErrorCode MatDestroy_MPIAIJ_SeqsToMPI(Mat A)
4540: {
4541: PetscErrorCode ierr;
4542: Mat_Merge_SeqsToMPI *merge;
4543: PetscContainer container;
4546: PetscObjectQuery((PetscObject)A,"MatMergeSeqsToMPI",(PetscObject*)&container);
4547: if (container) {
4548: PetscContainerGetPointer(container,(void**)&merge);
4549: PetscFree(merge->id_r);
4550: PetscFree(merge->len_s);
4551: PetscFree(merge->len_r);
4552: PetscFree(merge->bi);
4553: PetscFree(merge->bj);
4554: PetscFree(merge->buf_ri[0]);
4555: PetscFree(merge->buf_ri);
4556: PetscFree(merge->buf_rj[0]);
4557: PetscFree(merge->buf_rj);
4558: PetscFree(merge->coi);
4559: PetscFree(merge->coj);
4560: PetscFree(merge->owners_co);
4561: PetscLayoutDestroy(&merge->rowmap);
4562: PetscFree(merge);
4563: PetscObjectCompose((PetscObject)A,"MatMergeSeqsToMPI",0);
4564: }
4565: MatDestroy_MPIAIJ(A);
4566: return(0);
4567: }
4569: #include <../src/mat/utils/freespace.h>
4570: #include <petscbt.h>
4572: PetscErrorCode MatCreateMPIAIJSumSeqAIJNumeric(Mat seqmat,Mat mpimat)
4573: {
4574: PetscErrorCode ierr;
4575: MPI_Comm comm;
4576: Mat_SeqAIJ *a =(Mat_SeqAIJ*)seqmat->data;
4577: PetscMPIInt size,rank,taga,*len_s;
4578: PetscInt N=mpimat->cmap->N,i,j,*owners,*ai=a->i,*aj;
4579: PetscInt proc,m;
4580: PetscInt **buf_ri,**buf_rj;
4581: PetscInt k,anzi,*bj_i,*bi,*bj,arow,bnzi,nextaj;
4582: PetscInt nrows,**buf_ri_k,**nextrow,**nextai;
4583: MPI_Request *s_waits,*r_waits;
4584: MPI_Status *status;
4585: MatScalar *aa=a->a;
4586: MatScalar **abuf_r,*ba_i;
4587: Mat_Merge_SeqsToMPI *merge;
4588: PetscContainer container;
4591: PetscObjectGetComm((PetscObject)mpimat,&comm);
4592: PetscLogEventBegin(MAT_Seqstompinum,seqmat,0,0,0);
4594: MPI_Comm_size(comm,&size);
4595: MPI_Comm_rank(comm,&rank);
4597: PetscObjectQuery((PetscObject)mpimat,"MatMergeSeqsToMPI",(PetscObject*)&container);
4598: PetscContainerGetPointer(container,(void**)&merge);
4600: bi = merge->bi;
4601: bj = merge->bj;
4602: buf_ri = merge->buf_ri;
4603: buf_rj = merge->buf_rj;
4605: PetscMalloc1(size,&status);
4606: owners = merge->rowmap->range;
4607: len_s = merge->len_s;
4609: /* send and recv matrix values */
4610: /*-----------------------------*/
4611: PetscObjectGetNewTag((PetscObject)mpimat,&taga);
4612: PetscPostIrecvScalar(comm,taga,merge->nrecv,merge->id_r,merge->len_r,&abuf_r,&r_waits);
4614: PetscMalloc1(merge->nsend+1,&s_waits);
4615: for (proc=0,k=0; proc<size; proc++) {
4616: if (!len_s[proc]) continue;
4617: i = owners[proc];
4618: MPI_Isend(aa+ai[i],len_s[proc],MPIU_MATSCALAR,proc,taga,comm,s_waits+k);
4619: k++;
4620: }
4622: if (merge->nrecv) {MPI_Waitall(merge->nrecv,r_waits,status);}
4623: if (merge->nsend) {MPI_Waitall(merge->nsend,s_waits,status);}
4624: PetscFree(status);
4626: PetscFree(s_waits);
4627: PetscFree(r_waits);
4629: /* insert mat values of mpimat */
4630: /*----------------------------*/
4631: PetscMalloc1(N,&ba_i);
4632: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
4634: for (k=0; k<merge->nrecv; k++) {
4635: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4636: nrows = *(buf_ri_k[k]);
4637: nextrow[k] = buf_ri_k[k]+1; /* next row number of k-th recved i-structure */
4638: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
4639: }
4641: /* set values of ba */
4642: m = merge->rowmap->n;
4643: for (i=0; i<m; i++) {
4644: arow = owners[rank] + i;
4645: bj_i = bj+bi[i]; /* col indices of the i-th row of mpimat */
4646: bnzi = bi[i+1] - bi[i];
4647: PetscMemzero(ba_i,bnzi*sizeof(PetscScalar));
4649: /* add local non-zero vals of this proc's seqmat into ba */
4650: anzi = ai[arow+1] - ai[arow];
4651: aj = a->j + ai[arow];
4652: aa = a->a + ai[arow];
4653: nextaj = 0;
4654: for (j=0; nextaj<anzi; j++) {
4655: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4656: ba_i[j] += aa[nextaj++];
4657: }
4658: }
4660: /* add received vals into ba */
4661: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4662: /* i-th row */
4663: if (i == *nextrow[k]) {
4664: anzi = *(nextai[k]+1) - *nextai[k];
4665: aj = buf_rj[k] + *(nextai[k]);
4666: aa = abuf_r[k] + *(nextai[k]);
4667: nextaj = 0;
4668: for (j=0; nextaj<anzi; j++) {
4669: if (*(bj_i + j) == aj[nextaj]) { /* bcol == acol */
4670: ba_i[j] += aa[nextaj++];
4671: }
4672: }
4673: nextrow[k]++; nextai[k]++;
4674: }
4675: }
4676: MatSetValues(mpimat,1,&arow,bnzi,bj_i,ba_i,INSERT_VALUES);
4677: }
4678: MatAssemblyBegin(mpimat,MAT_FINAL_ASSEMBLY);
4679: MatAssemblyEnd(mpimat,MAT_FINAL_ASSEMBLY);
4681: PetscFree(abuf_r[0]);
4682: PetscFree(abuf_r);
4683: PetscFree(ba_i);
4684: PetscFree3(buf_ri_k,nextrow,nextai);
4685: PetscLogEventEnd(MAT_Seqstompinum,seqmat,0,0,0);
4686: return(0);
4687: }
4689: PetscErrorCode MatCreateMPIAIJSumSeqAIJSymbolic(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,Mat *mpimat)
4690: {
4691: PetscErrorCode ierr;
4692: Mat B_mpi;
4693: Mat_SeqAIJ *a=(Mat_SeqAIJ*)seqmat->data;
4694: PetscMPIInt size,rank,tagi,tagj,*len_s,*len_si,*len_ri;
4695: PetscInt **buf_rj,**buf_ri,**buf_ri_k;
4696: PetscInt M=seqmat->rmap->n,N=seqmat->cmap->n,i,*owners,*ai=a->i,*aj=a->j;
4697: PetscInt len,proc,*dnz,*onz,bs,cbs;
4698: PetscInt k,anzi,*bi,*bj,*lnk,nlnk,arow,bnzi,nspacedouble=0;
4699: PetscInt nrows,*buf_s,*buf_si,*buf_si_i,**nextrow,**nextai;
4700: MPI_Request *si_waits,*sj_waits,*ri_waits,*rj_waits;
4701: MPI_Status *status;
4702: PetscFreeSpaceList free_space=NULL,current_space=NULL;
4703: PetscBT lnkbt;
4704: Mat_Merge_SeqsToMPI *merge;
4705: PetscContainer container;
4708: PetscLogEventBegin(MAT_Seqstompisym,seqmat,0,0,0);
4710: /* make sure it is a PETSc comm */
4711: PetscCommDuplicate(comm,&comm,NULL);
4712: MPI_Comm_size(comm,&size);
4713: MPI_Comm_rank(comm,&rank);
4715: PetscNew(&merge);
4716: PetscMalloc1(size,&status);
4718: /* determine row ownership */
4719: /*---------------------------------------------------------*/
4720: PetscLayoutCreate(comm,&merge->rowmap);
4721: PetscLayoutSetLocalSize(merge->rowmap,m);
4722: PetscLayoutSetSize(merge->rowmap,M);
4723: PetscLayoutSetBlockSize(merge->rowmap,1);
4724: PetscLayoutSetUp(merge->rowmap);
4725: PetscMalloc1(size,&len_si);
4726: PetscMalloc1(size,&merge->len_s);
4728: m = merge->rowmap->n;
4729: owners = merge->rowmap->range;
4731: /* determine the number of messages to send, their lengths */
4732: /*---------------------------------------------------------*/
4733: len_s = merge->len_s;
4735: len = 0; /* length of buf_si[] */
4736: merge->nsend = 0;
4737: for (proc=0; proc<size; proc++) {
4738: len_si[proc] = 0;
4739: if (proc == rank) {
4740: len_s[proc] = 0;
4741: } else {
4742: len_si[proc] = owners[proc+1] - owners[proc] + 1;
4743: len_s[proc] = ai[owners[proc+1]] - ai[owners[proc]]; /* num of rows to be sent to [proc] */
4744: }
4745: if (len_s[proc]) {
4746: merge->nsend++;
4747: nrows = 0;
4748: for (i=owners[proc]; i<owners[proc+1]; i++) {
4749: if (ai[i+1] > ai[i]) nrows++;
4750: }
4751: len_si[proc] = 2*(nrows+1);
4752: len += len_si[proc];
4753: }
4754: }
4756: /* determine the number and length of messages to receive for ij-structure */
4757: /*-------------------------------------------------------------------------*/
4758: PetscGatherNumberOfMessages(comm,NULL,len_s,&merge->nrecv);
4759: PetscGatherMessageLengths2(comm,merge->nsend,merge->nrecv,len_s,len_si,&merge->id_r,&merge->len_r,&len_ri);
4761: /* post the Irecv of j-structure */
4762: /*-------------------------------*/
4763: PetscCommGetNewTag(comm,&tagj);
4764: PetscPostIrecvInt(comm,tagj,merge->nrecv,merge->id_r,merge->len_r,&buf_rj,&rj_waits);
4766: /* post the Isend of j-structure */
4767: /*--------------------------------*/
4768: PetscMalloc2(merge->nsend,&si_waits,merge->nsend,&sj_waits);
4770: for (proc=0, k=0; proc<size; proc++) {
4771: if (!len_s[proc]) continue;
4772: i = owners[proc];
4773: MPI_Isend(aj+ai[i],len_s[proc],MPIU_INT,proc,tagj,comm,sj_waits+k);
4774: k++;
4775: }
4777: /* receives and sends of j-structure are complete */
4778: /*------------------------------------------------*/
4779: if (merge->nrecv) {MPI_Waitall(merge->nrecv,rj_waits,status);}
4780: if (merge->nsend) {MPI_Waitall(merge->nsend,sj_waits,status);}
4782: /* send and recv i-structure */
4783: /*---------------------------*/
4784: PetscCommGetNewTag(comm,&tagi);
4785: PetscPostIrecvInt(comm,tagi,merge->nrecv,merge->id_r,len_ri,&buf_ri,&ri_waits);
4787: PetscMalloc1(len+1,&buf_s);
4788: buf_si = buf_s; /* points to the beginning of k-th msg to be sent */
4789: for (proc=0,k=0; proc<size; proc++) {
4790: if (!len_s[proc]) continue;
4791: /* form outgoing message for i-structure:
4792: buf_si[0]: nrows to be sent
4793: [1:nrows]: row index (global)
4794: [nrows+1:2*nrows+1]: i-structure index
4795: */
4796: /*-------------------------------------------*/
4797: nrows = len_si[proc]/2 - 1;
4798: buf_si_i = buf_si + nrows+1;
4799: buf_si[0] = nrows;
4800: buf_si_i[0] = 0;
4801: nrows = 0;
4802: for (i=owners[proc]; i<owners[proc+1]; i++) {
4803: anzi = ai[i+1] - ai[i];
4804: if (anzi) {
4805: buf_si_i[nrows+1] = buf_si_i[nrows] + anzi; /* i-structure */
4806: buf_si[nrows+1] = i-owners[proc]; /* local row index */
4807: nrows++;
4808: }
4809: }
4810: MPI_Isend(buf_si,len_si[proc],MPIU_INT,proc,tagi,comm,si_waits+k);
4811: k++;
4812: buf_si += len_si[proc];
4813: }
4815: if (merge->nrecv) {MPI_Waitall(merge->nrecv,ri_waits,status);}
4816: if (merge->nsend) {MPI_Waitall(merge->nsend,si_waits,status);}
4818: PetscInfo2(seqmat,"nsend: %D, nrecv: %D\n",merge->nsend,merge->nrecv);
4819: for (i=0; i<merge->nrecv; i++) {
4820: PetscInfo3(seqmat,"recv len_ri=%D, len_rj=%D from [%D]\n",len_ri[i],merge->len_r[i],merge->id_r[i]);
4821: }
4823: PetscFree(len_si);
4824: PetscFree(len_ri);
4825: PetscFree(rj_waits);
4826: PetscFree2(si_waits,sj_waits);
4827: PetscFree(ri_waits);
4828: PetscFree(buf_s);
4829: PetscFree(status);
4831: /* compute a local seq matrix in each processor */
4832: /*----------------------------------------------*/
4833: /* allocate bi array and free space for accumulating nonzero column info */
4834: PetscMalloc1(m+1,&bi);
4835: bi[0] = 0;
4837: /* create and initialize a linked list */
4838: nlnk = N+1;
4839: PetscLLCreate(N,N,nlnk,lnk,lnkbt);
4841: /* initial FreeSpace size is 2*(num of local nnz(seqmat)) */
4842: len = ai[owners[rank+1]] - ai[owners[rank]];
4843: PetscFreeSpaceGet(PetscIntMultTruncate(2,len)+1,&free_space);
4845: current_space = free_space;
4847: /* determine symbolic info for each local row */
4848: PetscMalloc3(merge->nrecv,&buf_ri_k,merge->nrecv,&nextrow,merge->nrecv,&nextai);
4850: for (k=0; k<merge->nrecv; k++) {
4851: buf_ri_k[k] = buf_ri[k]; /* beginning of k-th recved i-structure */
4852: nrows = *buf_ri_k[k];
4853: nextrow[k] = buf_ri_k[k] + 1; /* next row number of k-th recved i-structure */
4854: nextai[k] = buf_ri_k[k] + (nrows + 1); /* poins to the next i-structure of k-th recved i-structure */
4855: }
4857: MatPreallocateInitialize(comm,m,n,dnz,onz);
4858: len = 0;
4859: for (i=0; i<m; i++) {
4860: bnzi = 0;
4861: /* add local non-zero cols of this proc's seqmat into lnk */
4862: arow = owners[rank] + i;
4863: anzi = ai[arow+1] - ai[arow];
4864: aj = a->j + ai[arow];
4865: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4866: bnzi += nlnk;
4867: /* add received col data into lnk */
4868: for (k=0; k<merge->nrecv; k++) { /* k-th received message */
4869: if (i == *nextrow[k]) { /* i-th row */
4870: anzi = *(nextai[k]+1) - *nextai[k];
4871: aj = buf_rj[k] + *nextai[k];
4872: PetscLLAddSorted(anzi,aj,N,nlnk,lnk,lnkbt);
4873: bnzi += nlnk;
4874: nextrow[k]++; nextai[k]++;
4875: }
4876: }
4877: if (len < bnzi) len = bnzi; /* =max(bnzi) */
4879: /* if free space is not available, make more free space */
4880: if (current_space->local_remaining<bnzi) {
4881: PetscFreeSpaceGet(PetscIntSumTruncate(bnzi,current_space->total_array_size),¤t_space);
4882: nspacedouble++;
4883: }
4884: /* copy data into free space, then initialize lnk */
4885: PetscLLClean(N,N,bnzi,lnk,current_space->array,lnkbt);
4886: MatPreallocateSet(i+owners[rank],bnzi,current_space->array,dnz,onz);
4888: current_space->array += bnzi;
4889: current_space->local_used += bnzi;
4890: current_space->local_remaining -= bnzi;
4892: bi[i+1] = bi[i] + bnzi;
4893: }
4895: PetscFree3(buf_ri_k,nextrow,nextai);
4897: PetscMalloc1(bi[m]+1,&bj);
4898: PetscFreeSpaceContiguous(&free_space,bj);
4899: PetscLLDestroy(lnk,lnkbt);
4901: /* create symbolic parallel matrix B_mpi */
4902: /*---------------------------------------*/
4903: MatGetBlockSizes(seqmat,&bs,&cbs);
4904: MatCreate(comm,&B_mpi);
4905: if (n==PETSC_DECIDE) {
4906: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,N);
4907: } else {
4908: MatSetSizes(B_mpi,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
4909: }
4910: MatSetBlockSizes(B_mpi,bs,cbs);
4911: MatSetType(B_mpi,MATMPIAIJ);
4912: MatMPIAIJSetPreallocation(B_mpi,0,dnz,0,onz);
4913: MatPreallocateFinalize(dnz,onz);
4914: MatSetOption(B_mpi,MAT_NEW_NONZERO_ALLOCATION_ERR,PETSC_FALSE);
4916: /* B_mpi is not ready for use - assembly will be done by MatCreateMPIAIJSumSeqAIJNumeric() */
4917: B_mpi->assembled = PETSC_FALSE;
4918: B_mpi->ops->destroy = MatDestroy_MPIAIJ_SeqsToMPI;
4919: merge->bi = bi;
4920: merge->bj = bj;
4921: merge->buf_ri = buf_ri;
4922: merge->buf_rj = buf_rj;
4923: merge->coi = NULL;
4924: merge->coj = NULL;
4925: merge->owners_co = NULL;
4927: PetscCommDestroy(&comm);
4929: /* attach the supporting struct to B_mpi for reuse */
4930: PetscContainerCreate(PETSC_COMM_SELF,&container);
4931: PetscContainerSetPointer(container,merge);
4932: PetscObjectCompose((PetscObject)B_mpi,"MatMergeSeqsToMPI",(PetscObject)container);
4933: PetscContainerDestroy(&container);
4934: *mpimat = B_mpi;
4936: PetscLogEventEnd(MAT_Seqstompisym,seqmat,0,0,0);
4937: return(0);
4938: }
4940: /*@C
4941: MatCreateMPIAIJSumSeqAIJ - Creates a MATMPIAIJ matrix by adding sequential
4942: matrices from each processor
4944: Collective on MPI_Comm
4946: Input Parameters:
4947: + comm - the communicators the parallel matrix will live on
4948: . seqmat - the input sequential matrices
4949: . m - number of local rows (or PETSC_DECIDE)
4950: . n - number of local columns (or PETSC_DECIDE)
4951: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
4953: Output Parameter:
4954: . mpimat - the parallel matrix generated
4956: Level: advanced
4958: Notes:
4959: The dimensions of the sequential matrix in each processor MUST be the same.
4960: The input seqmat is included into the container "Mat_Merge_SeqsToMPI", and will be
4961: destroyed when mpimat is destroyed. Call PetscObjectQuery() to access seqmat.
4962: @*/
4963: PetscErrorCode MatCreateMPIAIJSumSeqAIJ(MPI_Comm comm,Mat seqmat,PetscInt m,PetscInt n,MatReuse scall,Mat *mpimat)
4964: {
4966: PetscMPIInt size;
4969: MPI_Comm_size(comm,&size);
4970: if (size == 1) {
4971: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4972: if (scall == MAT_INITIAL_MATRIX) {
4973: MatDuplicate(seqmat,MAT_COPY_VALUES,mpimat);
4974: } else {
4975: MatCopy(seqmat,*mpimat,SAME_NONZERO_PATTERN);
4976: }
4977: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4978: return(0);
4979: }
4980: PetscLogEventBegin(MAT_Seqstompi,seqmat,0,0,0);
4981: if (scall == MAT_INITIAL_MATRIX) {
4982: MatCreateMPIAIJSumSeqAIJSymbolic(comm,seqmat,m,n,mpimat);
4983: }
4984: MatCreateMPIAIJSumSeqAIJNumeric(seqmat,*mpimat);
4985: PetscLogEventEnd(MAT_Seqstompi,seqmat,0,0,0);
4986: return(0);
4987: }
4989: /*@
4990: MatMPIAIJGetLocalMat - Creates a SeqAIJ from a MATMPIAIJ matrix by taking all its local rows and putting them into a sequential matrix with
4991: mlocal rows and n columns. Where mlocal is the row count obtained with MatGetLocalSize() and n is the global column count obtained
4992: with MatGetSize()
4994: Not Collective
4996: Input Parameters:
4997: + A - the matrix
4998: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5000: Output Parameter:
5001: . A_loc - the local sequential matrix generated
5003: Level: developer
5005: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMatCondensed()
5007: @*/
5008: PetscErrorCode MatMPIAIJGetLocalMat(Mat A,MatReuse scall,Mat *A_loc)
5009: {
5011: Mat_MPIAIJ *mpimat=(Mat_MPIAIJ*)A->data;
5012: Mat_SeqAIJ *mat,*a,*b;
5013: PetscInt *ai,*aj,*bi,*bj,*cmap=mpimat->garray;
5014: MatScalar *aa,*ba,*cam;
5015: PetscScalar *ca;
5016: PetscInt am=A->rmap->n,i,j,k,cstart=A->cmap->rstart;
5017: PetscInt *ci,*cj,col,ncols_d,ncols_o,jo;
5018: PetscBool match;
5019: MPI_Comm comm;
5020: PetscMPIInt size;
5023: PetscStrbeginswith(((PetscObject)A)->type_name,MATMPIAIJ,&match);
5024: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5025: PetscObjectGetComm((PetscObject)A,&comm);
5026: MPI_Comm_size(comm,&size);
5027: if (size == 1 && scall == MAT_REUSE_MATRIX) return(0);
5029: PetscLogEventBegin(MAT_Getlocalmat,A,0,0,0);
5030: a = (Mat_SeqAIJ*)(mpimat->A)->data;
5031: b = (Mat_SeqAIJ*)(mpimat->B)->data;
5032: ai = a->i; aj = a->j; bi = b->i; bj = b->j;
5033: aa = a->a; ba = b->a;
5034: if (scall == MAT_INITIAL_MATRIX) {
5035: if (size == 1) {
5036: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ai,aj,aa,A_loc);
5037: return(0);
5038: }
5040: PetscMalloc1(1+am,&ci);
5041: ci[0] = 0;
5042: for (i=0; i<am; i++) {
5043: ci[i+1] = ci[i] + (ai[i+1] - ai[i]) + (bi[i+1] - bi[i]);
5044: }
5045: PetscMalloc1(1+ci[am],&cj);
5046: PetscMalloc1(1+ci[am],&ca);
5047: k = 0;
5048: for (i=0; i<am; i++) {
5049: ncols_o = bi[i+1] - bi[i];
5050: ncols_d = ai[i+1] - ai[i];
5051: /* off-diagonal portion of A */
5052: for (jo=0; jo<ncols_o; jo++) {
5053: col = cmap[*bj];
5054: if (col >= cstart) break;
5055: cj[k] = col; bj++;
5056: ca[k++] = *ba++;
5057: }
5058: /* diagonal portion of A */
5059: for (j=0; j<ncols_d; j++) {
5060: cj[k] = cstart + *aj++;
5061: ca[k++] = *aa++;
5062: }
5063: /* off-diagonal portion of A */
5064: for (j=jo; j<ncols_o; j++) {
5065: cj[k] = cmap[*bj++];
5066: ca[k++] = *ba++;
5067: }
5068: }
5069: /* put together the new matrix */
5070: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,am,A->cmap->N,ci,cj,ca,A_loc);
5071: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5072: /* Since these are PETSc arrays, change flags to free them as necessary. */
5073: mat = (Mat_SeqAIJ*)(*A_loc)->data;
5074: mat->free_a = PETSC_TRUE;
5075: mat->free_ij = PETSC_TRUE;
5076: mat->nonew = 0;
5077: } else if (scall == MAT_REUSE_MATRIX) {
5078: mat=(Mat_SeqAIJ*)(*A_loc)->data;
5079: ci = mat->i; cj = mat->j; cam = mat->a;
5080: for (i=0; i<am; i++) {
5081: /* off-diagonal portion of A */
5082: ncols_o = bi[i+1] - bi[i];
5083: for (jo=0; jo<ncols_o; jo++) {
5084: col = cmap[*bj];
5085: if (col >= cstart) break;
5086: *cam++ = *ba++; bj++;
5087: }
5088: /* diagonal portion of A */
5089: ncols_d = ai[i+1] - ai[i];
5090: for (j=0; j<ncols_d; j++) *cam++ = *aa++;
5091: /* off-diagonal portion of A */
5092: for (j=jo; j<ncols_o; j++) {
5093: *cam++ = *ba++; bj++;
5094: }
5095: }
5096: } else SETERRQ1(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONG,"Invalid MatReuse %d",(int)scall);
5097: PetscLogEventEnd(MAT_Getlocalmat,A,0,0,0);
5098: return(0);
5099: }
5101: /*@C
5102: MatMPIAIJGetLocalMatCondensed - Creates a SeqAIJ matrix from an MATMPIAIJ matrix by taking all its local rows and NON-ZERO columns
5104: Not Collective
5106: Input Parameters:
5107: + A - the matrix
5108: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5109: - row, col - index sets of rows and columns to extract (or NULL)
5111: Output Parameter:
5112: . A_loc - the local sequential matrix generated
5114: Level: developer
5116: .seealso: MatGetOwnershipRange(), MatMPIAIJGetLocalMat()
5118: @*/
5119: PetscErrorCode MatMPIAIJGetLocalMatCondensed(Mat A,MatReuse scall,IS *row,IS *col,Mat *A_loc)
5120: {
5121: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5123: PetscInt i,start,end,ncols,nzA,nzB,*cmap,imark,*idx;
5124: IS isrowa,iscola;
5125: Mat *aloc;
5126: PetscBool match;
5129: PetscObjectTypeCompare((PetscObject)A,MATMPIAIJ,&match);
5130: if (!match) SETERRQ(PetscObjectComm((PetscObject)A), PETSC_ERR_SUP,"Requires MATMPIAIJ matrix as input");
5131: PetscLogEventBegin(MAT_Getlocalmatcondensed,A,0,0,0);
5132: if (!row) {
5133: start = A->rmap->rstart; end = A->rmap->rend;
5134: ISCreateStride(PETSC_COMM_SELF,end-start,start,1,&isrowa);
5135: } else {
5136: isrowa = *row;
5137: }
5138: if (!col) {
5139: start = A->cmap->rstart;
5140: cmap = a->garray;
5141: nzA = a->A->cmap->n;
5142: nzB = a->B->cmap->n;
5143: PetscMalloc1(nzA+nzB, &idx);
5144: ncols = 0;
5145: for (i=0; i<nzB; i++) {
5146: if (cmap[i] < start) idx[ncols++] = cmap[i];
5147: else break;
5148: }
5149: imark = i;
5150: for (i=0; i<nzA; i++) idx[ncols++] = start + i;
5151: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i];
5152: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&iscola);
5153: } else {
5154: iscola = *col;
5155: }
5156: if (scall != MAT_INITIAL_MATRIX) {
5157: PetscMalloc1(1,&aloc);
5158: aloc[0] = *A_loc;
5159: }
5160: MatCreateSubMatrices(A,1,&isrowa,&iscola,scall,&aloc);
5161: if (!col) { /* attach global id of condensed columns */
5162: PetscObjectCompose((PetscObject)aloc[0],"_petsc_GetLocalMatCondensed_iscol",(PetscObject)iscola);
5163: }
5164: *A_loc = aloc[0];
5165: PetscFree(aloc);
5166: if (!row) {
5167: ISDestroy(&isrowa);
5168: }
5169: if (!col) {
5170: ISDestroy(&iscola);
5171: }
5172: PetscLogEventEnd(MAT_Getlocalmatcondensed,A,0,0,0);
5173: return(0);
5174: }
5176: /*@C
5177: MatGetBrowsOfAcols - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns of local A
5179: Collective on Mat
5181: Input Parameters:
5182: + A,B - the matrices in mpiaij format
5183: . scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5184: - rowb, colb - index sets of rows and columns of B to extract (or NULL)
5186: Output Parameter:
5187: + rowb, colb - index sets of rows and columns of B to extract
5188: - B_seq - the sequential matrix generated
5190: Level: developer
5192: @*/
5193: PetscErrorCode MatGetBrowsOfAcols(Mat A,Mat B,MatReuse scall,IS *rowb,IS *colb,Mat *B_seq)
5194: {
5195: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5197: PetscInt *idx,i,start,ncols,nzA,nzB,*cmap,imark;
5198: IS isrowb,iscolb;
5199: Mat *bseq=NULL;
5202: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5203: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%D, %D) != (%D,%D)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5204: }
5205: PetscLogEventBegin(MAT_GetBrowsOfAcols,A,B,0,0);
5207: if (scall == MAT_INITIAL_MATRIX) {
5208: start = A->cmap->rstart;
5209: cmap = a->garray;
5210: nzA = a->A->cmap->n;
5211: nzB = a->B->cmap->n;
5212: PetscMalloc1(nzA+nzB, &idx);
5213: ncols = 0;
5214: for (i=0; i<nzB; i++) { /* row < local row index */
5215: if (cmap[i] < start) idx[ncols++] = cmap[i];
5216: else break;
5217: }
5218: imark = i;
5219: for (i=0; i<nzA; i++) idx[ncols++] = start + i; /* local rows */
5220: for (i=imark; i<nzB; i++) idx[ncols++] = cmap[i]; /* row > local row index */
5221: ISCreateGeneral(PETSC_COMM_SELF,ncols,idx,PETSC_OWN_POINTER,&isrowb);
5222: ISCreateStride(PETSC_COMM_SELF,B->cmap->N,0,1,&iscolb);
5223: } else {
5224: if (!rowb || !colb) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_SUP,"IS rowb and colb must be provided for MAT_REUSE_MATRIX");
5225: isrowb = *rowb; iscolb = *colb;
5226: PetscMalloc1(1,&bseq);
5227: bseq[0] = *B_seq;
5228: }
5229: MatCreateSubMatrices(B,1,&isrowb,&iscolb,scall,&bseq);
5230: *B_seq = bseq[0];
5231: PetscFree(bseq);
5232: if (!rowb) {
5233: ISDestroy(&isrowb);
5234: } else {
5235: *rowb = isrowb;
5236: }
5237: if (!colb) {
5238: ISDestroy(&iscolb);
5239: } else {
5240: *colb = iscolb;
5241: }
5242: PetscLogEventEnd(MAT_GetBrowsOfAcols,A,B,0,0);
5243: return(0);
5244: }
5246: /*
5247: MatGetBrowsOfAoCols_MPIAIJ - Creates a SeqAIJ matrix by taking rows of B that equal to nonzero columns
5248: of the OFF-DIAGONAL portion of local A
5250: Collective on Mat
5252: Input Parameters:
5253: + A,B - the matrices in mpiaij format
5254: - scall - either MAT_INITIAL_MATRIX or MAT_REUSE_MATRIX
5256: Output Parameter:
5257: + startsj_s - starting point in B's sending j-arrays, saved for MAT_REUSE (or NULL)
5258: . startsj_r - starting point in B's receiving j-arrays, saved for MAT_REUSE (or NULL)
5259: . bufa_ptr - array for sending matrix values, saved for MAT_REUSE (or NULL)
5260: - B_oth - the sequential matrix generated with size aBn=a->B->cmap->n by B->cmap->N
5262: Developer Notes: This directly accesses information inside the VecScatter associated with the matrix-vector product
5263: for this matrix. This is not desirable..
5265: Level: developer
5267: */
5268: PetscErrorCode MatGetBrowsOfAoCols_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscInt **startsj_s,PetscInt **startsj_r,MatScalar **bufa_ptr,Mat *B_oth)
5269: {
5270: PetscErrorCode ierr;
5271: Mat_MPIAIJ *a=(Mat_MPIAIJ*)A->data;
5272: Mat_SeqAIJ *b_oth;
5273: VecScatter ctx;
5274: MPI_Comm comm;
5275: const PetscMPIInt *rprocs,*sprocs;
5276: const PetscInt *srow,*rstarts,*sstarts;
5277: PetscInt *rowlen,*bufj,*bufJ,ncols,aBn=a->B->cmap->n,row,*b_othi,*b_othj,*rvalues=NULL,*svalues=NULL,*cols,sbs,rbs;
5278: PetscInt i,j,k=0,l,ll,nrecvs,nsends,nrows,*rstartsj = 0,*sstartsj,len;
5279: PetscScalar *b_otha,*bufa,*bufA,*vals;
5280: MPI_Request *rwaits = NULL,*swaits = NULL;
5281: MPI_Status rstatus;
5282: PetscMPIInt jj,size,tag,rank,nsends_mpi,nrecvs_mpi;
5285: PetscObjectGetComm((PetscObject)A,&comm);
5286: MPI_Comm_size(comm,&size);
5288: if (A->cmap->rstart != B->rmap->rstart || A->cmap->rend != B->rmap->rend) {
5289: SETERRQ4(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"Matrix local dimensions are incompatible, (%d, %d) != (%d,%d)",A->cmap->rstart,A->cmap->rend,B->rmap->rstart,B->rmap->rend);
5290: }
5291: PetscLogEventBegin(MAT_GetBrowsOfAocols,A,B,0,0);
5292: MPI_Comm_rank(comm,&rank);
5294: if (size == 1) {
5295: startsj_s = NULL;
5296: bufa_ptr = NULL;
5297: *B_oth = NULL;
5298: return(0);
5299: }
5301: ctx = a->Mvctx;
5302: tag = ((PetscObject)ctx)->tag;
5304: if (ctx->inuse) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE," Scatter ctx already in use");
5305: VecScatterGetRemote_Private(ctx,PETSC_TRUE/*send*/,&nsends,&sstarts,&srow,&sprocs,&sbs);
5306: /* rprocs[] must be ordered so that indices received from them are ordered in rvalues[], which is key to algorithms used in this subroutine */
5307: VecScatterGetRemoteOrdered_Private(ctx,PETSC_FALSE/*recv*/,&nrecvs,&rstarts,NULL/*indices not needed*/,&rprocs,&rbs);
5308: PetscMPIIntCast(nsends,&nsends_mpi);
5309: PetscMPIIntCast(nrecvs,&nrecvs_mpi);
5310: PetscMalloc2(nrecvs,&rwaits,nsends,&swaits);
5312: if (!startsj_s || !bufa_ptr) scall = MAT_INITIAL_MATRIX;
5313: if (scall == MAT_INITIAL_MATRIX) {
5314: /* i-array */
5315: /*---------*/
5316: /* post receives */
5317: if (nrecvs) {PetscMalloc1(rbs*(rstarts[nrecvs] - rstarts[0]),&rvalues);} /* rstarts can be NULL when nrecvs=0 */
5318: for (i=0; i<nrecvs; i++) {
5319: rowlen = rvalues + rstarts[i]*rbs;
5320: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of indices to be received */
5321: MPI_Irecv(rowlen,nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5322: }
5324: /* pack the outgoing message */
5325: PetscMalloc2(nsends+1,&sstartsj,nrecvs+1,&rstartsj);
5327: sstartsj[0] = 0;
5328: rstartsj[0] = 0;
5329: len = 0; /* total length of j or a array to be sent */
5330: if (nsends) {
5331: k = sstarts[0]; /* ATTENTION: sstarts[0] and rstarts[0] are not necessarily zero */
5332: PetscMalloc1(sbs*(sstarts[nsends]-sstarts[0]),&svalues);
5333: }
5334: for (i=0; i<nsends; i++) {
5335: rowlen = svalues + (sstarts[i]-sstarts[0])*sbs;
5336: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5337: for (j=0; j<nrows; j++) {
5338: row = srow[k] + B->rmap->range[rank]; /* global row idx */
5339: for (l=0; l<sbs; l++) {
5340: MatGetRow_MPIAIJ(B,row+l,&ncols,NULL,NULL); /* rowlength */
5342: rowlen[j*sbs+l] = ncols;
5344: len += ncols;
5345: MatRestoreRow_MPIAIJ(B,row+l,&ncols,NULL,NULL);
5346: }
5347: k++;
5348: }
5349: MPI_Isend(rowlen,nrows*sbs,MPIU_INT,sprocs[i],tag,comm,swaits+i);
5351: sstartsj[i+1] = len; /* starting point of (i+1)-th outgoing msg in bufj and bufa */
5352: }
5353: /* recvs and sends of i-array are completed */
5354: i = nrecvs;
5355: while (i--) {
5356: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5357: }
5358: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5359: PetscFree(svalues);
5361: /* allocate buffers for sending j and a arrays */
5362: PetscMalloc1(len+1,&bufj);
5363: PetscMalloc1(len+1,&bufa);
5365: /* create i-array of B_oth */
5366: PetscMalloc1(aBn+2,&b_othi);
5368: b_othi[0] = 0;
5369: len = 0; /* total length of j or a array to be received */
5370: k = 0;
5371: for (i=0; i<nrecvs; i++) {
5372: rowlen = rvalues + (rstarts[i]-rstarts[0])*rbs;
5373: nrows = (rstarts[i+1]-rstarts[i])*rbs; /* num of rows to be received */
5374: for (j=0; j<nrows; j++) {
5375: b_othi[k+1] = b_othi[k] + rowlen[j];
5376: PetscIntSumError(rowlen[j],len,&len);
5377: k++;
5378: }
5379: rstartsj[i+1] = len; /* starting point of (i+1)-th incoming msg in bufj and bufa */
5380: }
5381: PetscFree(rvalues);
5383: /* allocate space for j and a arrrays of B_oth */
5384: PetscMalloc1(b_othi[aBn]+1,&b_othj);
5385: PetscMalloc1(b_othi[aBn]+1,&b_otha);
5387: /* j-array */
5388: /*---------*/
5389: /* post receives of j-array */
5390: for (i=0; i<nrecvs; i++) {
5391: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5392: MPI_Irecv(b_othj+rstartsj[i],nrows,MPIU_INT,rprocs[i],tag,comm,rwaits+i);
5393: }
5395: /* pack the outgoing message j-array */
5396: if (nsends) k = sstarts[0];
5397: for (i=0; i<nsends; i++) {
5398: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5399: bufJ = bufj+sstartsj[i];
5400: for (j=0; j<nrows; j++) {
5401: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5402: for (ll=0; ll<sbs; ll++) {
5403: MatGetRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5404: for (l=0; l<ncols; l++) {
5405: *bufJ++ = cols[l];
5406: }
5407: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,&cols,NULL);
5408: }
5409: }
5410: MPI_Isend(bufj+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_INT,sprocs[i],tag,comm,swaits+i);
5411: }
5413: /* recvs and sends of j-array are completed */
5414: i = nrecvs;
5415: while (i--) {
5416: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5417: }
5418: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5419: } else if (scall == MAT_REUSE_MATRIX) {
5420: sstartsj = *startsj_s;
5421: rstartsj = *startsj_r;
5422: bufa = *bufa_ptr;
5423: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5424: b_otha = b_oth->a;
5425: } else SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE, "Matrix P does not posses an object container");
5427: /* a-array */
5428: /*---------*/
5429: /* post receives of a-array */
5430: for (i=0; i<nrecvs; i++) {
5431: nrows = rstartsj[i+1]-rstartsj[i]; /* length of the msg received */
5432: MPI_Irecv(b_otha+rstartsj[i],nrows,MPIU_SCALAR,rprocs[i],tag,comm,rwaits+i);
5433: }
5435: /* pack the outgoing message a-array */
5436: if (nsends) k = sstarts[0];
5437: for (i=0; i<nsends; i++) {
5438: nrows = sstarts[i+1]-sstarts[i]; /* num of block rows */
5439: bufA = bufa+sstartsj[i];
5440: for (j=0; j<nrows; j++) {
5441: row = srow[k++] + B->rmap->range[rank]; /* global row idx */
5442: for (ll=0; ll<sbs; ll++) {
5443: MatGetRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5444: for (l=0; l<ncols; l++) {
5445: *bufA++ = vals[l];
5446: }
5447: MatRestoreRow_MPIAIJ(B,row+ll,&ncols,NULL,&vals);
5448: }
5449: }
5450: MPI_Isend(bufa+sstartsj[i],sstartsj[i+1]-sstartsj[i],MPIU_SCALAR,sprocs[i],tag,comm,swaits+i);
5451: }
5452: /* recvs and sends of a-array are completed */
5453: i = nrecvs;
5454: while (i--) {
5455: MPI_Waitany(nrecvs_mpi,rwaits,&jj,&rstatus);
5456: }
5457: if (nsends) {MPI_Waitall(nsends_mpi,swaits,MPI_STATUSES_IGNORE);}
5458: PetscFree2(rwaits,swaits);
5460: if (scall == MAT_INITIAL_MATRIX) {
5461: /* put together the new matrix */
5462: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,aBn,B->cmap->N,b_othi,b_othj,b_otha,B_oth);
5464: /* MatCreateSeqAIJWithArrays flags matrix so PETSc doesn't free the user's arrays. */
5465: /* Since these are PETSc arrays, change flags to free them as necessary. */
5466: b_oth = (Mat_SeqAIJ*)(*B_oth)->data;
5467: b_oth->free_a = PETSC_TRUE;
5468: b_oth->free_ij = PETSC_TRUE;
5469: b_oth->nonew = 0;
5471: PetscFree(bufj);
5472: if (!startsj_s || !bufa_ptr) {
5473: PetscFree2(sstartsj,rstartsj);
5474: PetscFree(bufa_ptr);
5475: } else {
5476: *startsj_s = sstartsj;
5477: *startsj_r = rstartsj;
5478: *bufa_ptr = bufa;
5479: }
5480: }
5482: VecScatterRestoreRemote_Private(ctx,PETSC_TRUE,&nsends,&sstarts,&srow,&sprocs,&sbs);
5483: VecScatterRestoreRemoteOrdered_Private(ctx,PETSC_FALSE,&nrecvs,&rstarts,NULL,&rprocs,&rbs);
5484: PetscLogEventEnd(MAT_GetBrowsOfAocols,A,B,0,0);
5485: return(0);
5486: }
5488: /*@C
5489: MatGetCommunicationStructs - Provides access to the communication structures used in matrix-vector multiplication.
5491: Not Collective
5493: Input Parameters:
5494: . A - The matrix in mpiaij format
5496: Output Parameter:
5497: + lvec - The local vector holding off-process values from the argument to a matrix-vector product
5498: . colmap - A map from global column index to local index into lvec
5499: - multScatter - A scatter from the argument of a matrix-vector product to lvec
5501: Level: developer
5503: @*/
5504: #if defined(PETSC_USE_CTABLE)
5505: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscTable *colmap, VecScatter *multScatter)
5506: #else
5507: PetscErrorCode MatGetCommunicationStructs(Mat A, Vec *lvec, PetscInt *colmap[], VecScatter *multScatter)
5508: #endif
5509: {
5510: Mat_MPIAIJ *a;
5517: a = (Mat_MPIAIJ*) A->data;
5518: if (lvec) *lvec = a->lvec;
5519: if (colmap) *colmap = a->colmap;
5520: if (multScatter) *multScatter = a->Mvctx;
5521: return(0);
5522: }
5524: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJCRL(Mat,MatType,MatReuse,Mat*);
5525: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJPERM(Mat,MatType,MatReuse,Mat*);
5526: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJSELL(Mat,MatType,MatReuse,Mat*);
5527: #if defined(PETSC_HAVE_MKL_SPARSE)
5528: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPIAIJMKL(Mat,MatType,MatReuse,Mat*);
5529: #endif
5530: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISBAIJ(Mat,MatType,MatReuse,Mat*);
5531: #if defined(PETSC_HAVE_ELEMENTAL)
5532: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_Elemental(Mat,MatType,MatReuse,Mat*);
5533: #endif
5534: #if defined(PETSC_HAVE_HYPRE)
5535: PETSC_INTERN PetscErrorCode MatConvert_AIJ_HYPRE(Mat,MatType,MatReuse,Mat*);
5536: PETSC_INTERN PetscErrorCode MatMatMatMult_Transpose_AIJ_AIJ(Mat,Mat,Mat,MatReuse,PetscReal,Mat*);
5537: #endif
5538: PETSC_INTERN PetscErrorCode MatConvert_XAIJ_IS(Mat,MatType,MatReuse,Mat*);
5539: PETSC_INTERN PetscErrorCode MatConvert_MPIAIJ_MPISELL(Mat,MatType,MatReuse,Mat*);
5540: PETSC_INTERN PetscErrorCode MatPtAP_IS_XAIJ(Mat,Mat,MatReuse,PetscReal,Mat*);
5542: /*
5543: Computes (B'*A')' since computing B*A directly is untenable
5545: n p p
5546: ( ) ( ) ( )
5547: m ( A ) * n ( B ) = m ( C )
5548: ( ) ( ) ( )
5550: */
5551: PetscErrorCode MatMatMultNumeric_MPIDense_MPIAIJ(Mat A,Mat B,Mat C)
5552: {
5554: Mat At,Bt,Ct;
5557: MatTranspose(A,MAT_INITIAL_MATRIX,&At);
5558: MatTranspose(B,MAT_INITIAL_MATRIX,&Bt);
5559: MatMatMult(Bt,At,MAT_INITIAL_MATRIX,1.0,&Ct);
5560: MatDestroy(&At);
5561: MatDestroy(&Bt);
5562: MatTranspose(Ct,MAT_REUSE_MATRIX,&C);
5563: MatDestroy(&Ct);
5564: return(0);
5565: }
5567: PetscErrorCode MatMatMultSymbolic_MPIDense_MPIAIJ(Mat A,Mat B,PetscReal fill,Mat *C)
5568: {
5570: PetscInt m=A->rmap->n,n=B->cmap->n;
5571: Mat Cmat;
5574: if (A->cmap->n != B->rmap->n) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_SIZ,"A->cmap->n %d != B->rmap->n %d\n",A->cmap->n,B->rmap->n);
5575: MatCreate(PetscObjectComm((PetscObject)A),&Cmat);
5576: MatSetSizes(Cmat,m,n,PETSC_DETERMINE,PETSC_DETERMINE);
5577: MatSetBlockSizesFromMats(Cmat,A,B);
5578: MatSetType(Cmat,MATMPIDENSE);
5579: MatMPIDenseSetPreallocation(Cmat,NULL);
5580: MatAssemblyBegin(Cmat,MAT_FINAL_ASSEMBLY);
5581: MatAssemblyEnd(Cmat,MAT_FINAL_ASSEMBLY);
5583: Cmat->ops->matmultnumeric = MatMatMultNumeric_MPIDense_MPIAIJ;
5585: *C = Cmat;
5586: return(0);
5587: }
5589: /* ----------------------------------------------------------------*/
5590: PETSC_INTERN PetscErrorCode MatMatMult_MPIDense_MPIAIJ(Mat A,Mat B,MatReuse scall,PetscReal fill,Mat *C)
5591: {
5595: if (scall == MAT_INITIAL_MATRIX) {
5596: PetscLogEventBegin(MAT_MatMultSymbolic,A,B,0,0);
5597: MatMatMultSymbolic_MPIDense_MPIAIJ(A,B,fill,C);
5598: PetscLogEventEnd(MAT_MatMultSymbolic,A,B,0,0);
5599: }
5600: PetscLogEventBegin(MAT_MatMultNumeric,A,B,0,0);
5601: MatMatMultNumeric_MPIDense_MPIAIJ(A,B,*C);
5602: PetscLogEventEnd(MAT_MatMultNumeric,A,B,0,0);
5603: return(0);
5604: }
5606: /*MC
5607: MATMPIAIJ - MATMPIAIJ = "mpiaij" - A matrix type to be used for parallel sparse matrices.
5609: Options Database Keys:
5610: . -mat_type mpiaij - sets the matrix type to "mpiaij" during a call to MatSetFromOptions()
5612: Level: beginner
5614: .seealso: MatCreateAIJ()
5615: M*/
5617: PETSC_EXTERN PetscErrorCode MatCreate_MPIAIJ(Mat B)
5618: {
5619: Mat_MPIAIJ *b;
5621: PetscMPIInt size;
5624: MPI_Comm_size(PetscObjectComm((PetscObject)B),&size);
5626: PetscNewLog(B,&b);
5627: B->data = (void*)b;
5628: PetscMemcpy(B->ops,&MatOps_Values,sizeof(struct _MatOps));
5629: B->assembled = PETSC_FALSE;
5630: B->insertmode = NOT_SET_VALUES;
5631: b->size = size;
5633: MPI_Comm_rank(PetscObjectComm((PetscObject)B),&b->rank);
5635: /* build cache for off array entries formed */
5636: MatStashCreate_Private(PetscObjectComm((PetscObject)B),1,&B->stash);
5638: b->donotstash = PETSC_FALSE;
5639: b->colmap = 0;
5640: b->garray = 0;
5641: b->roworiented = PETSC_TRUE;
5643: /* stuff used for matrix vector multiply */
5644: b->lvec = NULL;
5645: b->Mvctx = NULL;
5647: /* stuff for MatGetRow() */
5648: b->rowindices = 0;
5649: b->rowvalues = 0;
5650: b->getrowactive = PETSC_FALSE;
5652: /* flexible pointer used in CUSP/CUSPARSE classes */
5653: b->spptr = NULL;
5655: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetUseScalableIncreaseOverlap_C",MatMPIAIJSetUseScalableIncreaseOverlap_MPIAIJ);
5656: PetscObjectComposeFunction((PetscObject)B,"MatStoreValues_C",MatStoreValues_MPIAIJ);
5657: PetscObjectComposeFunction((PetscObject)B,"MatRetrieveValues_C",MatRetrieveValues_MPIAIJ);
5658: PetscObjectComposeFunction((PetscObject)B,"MatIsTranspose_C",MatIsTranspose_MPIAIJ);
5659: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocation_C",MatMPIAIJSetPreallocation_MPIAIJ);
5660: PetscObjectComposeFunction((PetscObject)B,"MatResetPreallocation_C",MatResetPreallocation_MPIAIJ);
5661: PetscObjectComposeFunction((PetscObject)B,"MatMPIAIJSetPreallocationCSR_C",MatMPIAIJSetPreallocationCSR_MPIAIJ);
5662: PetscObjectComposeFunction((PetscObject)B,"MatDiagonalScaleLocal_C",MatDiagonalScaleLocal_MPIAIJ);
5663: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijperm_C",MatConvert_MPIAIJ_MPIAIJPERM);
5664: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijsell_C",MatConvert_MPIAIJ_MPIAIJSELL);
5665: #if defined(PETSC_HAVE_MKL_SPARSE)
5666: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijmkl_C",MatConvert_MPIAIJ_MPIAIJMKL);
5667: #endif
5668: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpiaijcrl_C",MatConvert_MPIAIJ_MPIAIJCRL);
5669: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisbaij_C",MatConvert_MPIAIJ_MPISBAIJ);
5670: #if defined(PETSC_HAVE_ELEMENTAL)
5671: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_elemental_C",MatConvert_MPIAIJ_Elemental);
5672: #endif
5673: #if defined(PETSC_HAVE_HYPRE)
5674: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_hypre_C",MatConvert_AIJ_HYPRE);
5675: #endif
5676: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_is_C",MatConvert_XAIJ_IS);
5677: PetscObjectComposeFunction((PetscObject)B,"MatConvert_mpiaij_mpisell_C",MatConvert_MPIAIJ_MPISELL);
5678: PetscObjectComposeFunction((PetscObject)B,"MatMatMult_mpidense_mpiaij_C",MatMatMult_MPIDense_MPIAIJ);
5679: PetscObjectComposeFunction((PetscObject)B,"MatMatMultSymbolic_mpidense_mpiaij_C",MatMatMultSymbolic_MPIDense_MPIAIJ);
5680: PetscObjectComposeFunction((PetscObject)B,"MatMatMultNumeric_mpidense_mpiaij_C",MatMatMultNumeric_MPIDense_MPIAIJ);
5681: #if defined(PETSC_HAVE_HYPRE)
5682: PetscObjectComposeFunction((PetscObject)B,"MatMatMatMult_transpose_mpiaij_mpiaij_C",MatMatMatMult_Transpose_AIJ_AIJ);
5683: #endif
5684: PetscObjectComposeFunction((PetscObject)B,"MatPtAP_is_mpiaij_C",MatPtAP_IS_XAIJ);
5685: PetscObjectChangeTypeName((PetscObject)B,MATMPIAIJ);
5686: return(0);
5687: }
5689: /*@C
5690: MatCreateMPIAIJWithSplitArrays - creates a MPI AIJ matrix using arrays that contain the "diagonal"
5691: and "off-diagonal" part of the matrix in CSR format.
5693: Collective on MPI_Comm
5695: Input Parameters:
5696: + comm - MPI communicator
5697: . m - number of local rows (Cannot be PETSC_DECIDE)
5698: . n - This value should be the same as the local size used in creating the
5699: x vector for the matrix-vector product y = Ax. (or PETSC_DECIDE to have
5700: calculated if N is given) For square matrices n is almost always m.
5701: . M - number of global rows (or PETSC_DETERMINE to have calculated if m is given)
5702: . N - number of global columns (or PETSC_DETERMINE to have calculated if n is given)
5703: . i - row indices for "diagonal" portion of matrix; that is i[0] = 0, i[row] = i[row-1] + number of elements in that row of the matrix
5704: . j - column indices
5705: . a - matrix values
5706: . oi - row indices for "off-diagonal" portion of matrix; that is oi[0] = 0, oi[row] = oi[row-1] + number of elements in that row of the matrix
5707: . oj - column indices
5708: - oa - matrix values
5710: Output Parameter:
5711: . mat - the matrix
5713: Level: advanced
5715: Notes:
5716: The i, j, and a arrays ARE NOT copied by this routine into the internal format used by PETSc. The user
5717: must free the arrays once the matrix has been destroyed and not before.
5719: The i and j indices are 0 based
5721: See MatCreateAIJ() for the definition of "diagonal" and "off-diagonal" portion of the matrix
5723: This sets local rows and cannot be used to set off-processor values.
5725: Use of this routine is discouraged because it is inflexible and cumbersome to use. It is extremely rare that a
5726: legacy application natively assembles into exactly this split format. The code to do so is nontrivial and does
5727: not easily support in-place reassembly. It is recommended to use MatSetValues() (or a variant thereof) because
5728: the resulting assembly is easier to implement, will work with any matrix format, and the user does not have to
5729: keep track of the underlying array. Use MatSetOption(A,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE) to disable all
5730: communication if it is known that only local entries will be set.
5732: .keywords: matrix, aij, compressed row, sparse, parallel
5734: .seealso: MatCreate(), MatCreateSeqAIJ(), MatSetValues(), MatMPIAIJSetPreallocation(), MatMPIAIJSetPreallocationCSR(),
5735: MATMPIAIJ, MatCreateAIJ(), MatCreateMPIAIJWithArrays()
5736: @*/
5737: PetscErrorCode MatCreateMPIAIJWithSplitArrays(MPI_Comm comm,PetscInt m,PetscInt n,PetscInt M,PetscInt N,PetscInt i[],PetscInt j[],PetscScalar a[],PetscInt oi[], PetscInt oj[],PetscScalar oa[],Mat *mat)
5738: {
5740: Mat_MPIAIJ *maij;
5743: if (m < 0) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"local number of rows (m) cannot be PETSC_DECIDE, or negative");
5744: if (i[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"i (row indices) must start with 0");
5745: if (oi[0]) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"oi (row indices) must start with 0");
5746: MatCreate(comm,mat);
5747: MatSetSizes(*mat,m,n,M,N);
5748: MatSetType(*mat,MATMPIAIJ);
5749: maij = (Mat_MPIAIJ*) (*mat)->data;
5751: (*mat)->preallocated = PETSC_TRUE;
5753: PetscLayoutSetUp((*mat)->rmap);
5754: PetscLayoutSetUp((*mat)->cmap);
5756: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,n,i,j,a,&maij->A);
5757: MatCreateSeqAIJWithArrays(PETSC_COMM_SELF,m,(*mat)->cmap->N,oi,oj,oa,&maij->B);
5759: MatAssemblyBegin(maij->A,MAT_FINAL_ASSEMBLY);
5760: MatAssemblyEnd(maij->A,MAT_FINAL_ASSEMBLY);
5761: MatAssemblyBegin(maij->B,MAT_FINAL_ASSEMBLY);
5762: MatAssemblyEnd(maij->B,MAT_FINAL_ASSEMBLY);
5764: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_TRUE);
5765: MatAssemblyBegin(*mat,MAT_FINAL_ASSEMBLY);
5766: MatAssemblyEnd(*mat,MAT_FINAL_ASSEMBLY);
5767: MatSetOption(*mat,MAT_NO_OFF_PROC_ENTRIES,PETSC_FALSE);
5768: MatSetOption(*mat,MAT_NEW_NONZERO_LOCATION_ERR,PETSC_TRUE);
5769: return(0);
5770: }
5772: /*
5773: Special version for direct calls from Fortran
5774: */
5775: #include <petsc/private/fortranimpl.h>
5777: /* Change these macros so can be used in void function */
5778: #undef CHKERRQ
5779: #define CHKERRQ(ierr) CHKERRABORT(PETSC_COMM_WORLD,ierr)
5780: #undef SETERRQ2
5781: #define SETERRQ2(comm,ierr,b,c,d) CHKERRABORT(comm,ierr)
5782: #undef SETERRQ3
5783: #define SETERRQ3(comm,ierr,b,c,d,e) CHKERRABORT(comm,ierr)
5784: #undef SETERRQ
5785: #define SETERRQ(c,ierr,b) CHKERRABORT(c,ierr)
5787: #if defined(PETSC_HAVE_FORTRAN_CAPS)
5788: #define matsetvaluesmpiaij_ MATSETVALUESMPIAIJ
5789: #elif !defined(PETSC_HAVE_FORTRAN_UNDERSCORE)
5790: #define matsetvaluesmpiaij_ matsetvaluesmpiaij
5791: #else
5792: #endif
5793: PETSC_EXTERN void PETSC_STDCALL matsetvaluesmpiaij_(Mat *mmat,PetscInt *mm,const PetscInt im[],PetscInt *mn,const PetscInt in[],const PetscScalar v[],InsertMode *maddv,PetscErrorCode *_ierr)
5794: {
5795: Mat mat = *mmat;
5796: PetscInt m = *mm, n = *mn;
5797: InsertMode addv = *maddv;
5798: Mat_MPIAIJ *aij = (Mat_MPIAIJ*)mat->data;
5799: PetscScalar value;
5802: MatCheckPreallocated(mat,1);
5803: if (mat->insertmode == NOT_SET_VALUES) mat->insertmode = addv;
5805: #if defined(PETSC_USE_DEBUG)
5806: else if (mat->insertmode != addv) SETERRQ(PETSC_COMM_SELF,PETSC_ERR_ARG_WRONGSTATE,"Cannot mix add values and insert values");
5807: #endif
5808: {
5809: PetscInt i,j,rstart = mat->rmap->rstart,rend = mat->rmap->rend;
5810: PetscInt cstart = mat->cmap->rstart,cend = mat->cmap->rend,row,col;
5811: PetscBool roworiented = aij->roworiented;
5813: /* Some Variables required in the macro */
5814: Mat A = aij->A;
5815: Mat_SeqAIJ *a = (Mat_SeqAIJ*)A->data;
5816: PetscInt *aimax = a->imax,*ai = a->i,*ailen = a->ilen,*aj = a->j;
5817: MatScalar *aa = a->a;
5818: PetscBool ignorezeroentries = (((a->ignorezeroentries)&&(addv==ADD_VALUES)) ? PETSC_TRUE : PETSC_FALSE);
5819: Mat B = aij->B;
5820: Mat_SeqAIJ *b = (Mat_SeqAIJ*)B->data;
5821: PetscInt *bimax = b->imax,*bi = b->i,*bilen = b->ilen,*bj = b->j,bm = aij->B->rmap->n,am = aij->A->rmap->n;
5822: MatScalar *ba = b->a;
5824: PetscInt *rp1,*rp2,ii,nrow1,nrow2,_i,rmax1,rmax2,N,low1,high1,low2,high2,t,lastcol1,lastcol2;
5825: PetscInt nonew = a->nonew;
5826: MatScalar *ap1,*ap2;
5829: for (i=0; i<m; i++) {
5830: if (im[i] < 0) continue;
5831: #if defined(PETSC_USE_DEBUG)
5832: if (im[i] >= mat->rmap->N) SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Row too large: row %D max %D",im[i],mat->rmap->N-1);
5833: #endif
5834: if (im[i] >= rstart && im[i] < rend) {
5835: row = im[i] - rstart;
5836: lastcol1 = -1;
5837: rp1 = aj + ai[row];
5838: ap1 = aa + ai[row];
5839: rmax1 = aimax[row];
5840: nrow1 = ailen[row];
5841: low1 = 0;
5842: high1 = nrow1;
5843: lastcol2 = -1;
5844: rp2 = bj + bi[row];
5845: ap2 = ba + bi[row];
5846: rmax2 = bimax[row];
5847: nrow2 = bilen[row];
5848: low2 = 0;
5849: high2 = nrow2;
5851: for (j=0; j<n; j++) {
5852: if (roworiented) value = v[i*n+j];
5853: else value = v[i+j*m];
5854: if (in[j] >= cstart && in[j] < cend) {
5855: col = in[j] - cstart;
5856: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue;
5857: MatSetValues_SeqAIJ_A_Private(row,col,value,addv,im[i],in[j]);
5858: } else if (in[j] < 0) continue;
5859: #if defined(PETSC_USE_DEBUG)
5860: /* extra brace on SETERRQ2() is required for --with-errorchecking=0 - due to the next 'else' clause */
5861: else if (in[j] >= mat->cmap->N) {SETERRQ2(PETSC_COMM_SELF,PETSC_ERR_ARG_OUTOFRANGE,"Column too large: col %D max %D",in[j],mat->cmap->N-1);}
5862: #endif
5863: else {
5864: if (mat->was_assembled) {
5865: if (!aij->colmap) {
5866: MatCreateColmap_MPIAIJ_Private(mat);
5867: }
5868: #if defined(PETSC_USE_CTABLE)
5869: PetscTableFind(aij->colmap,in[j]+1,&col);
5870: col--;
5871: #else
5872: col = aij->colmap[in[j]] - 1;
5873: #endif
5874: if (ignorezeroentries && value == 0.0 && (addv == ADD_VALUES) && row != col) continue;
5875: if (col < 0 && !((Mat_SeqAIJ*)(aij->A->data))->nonew) {
5876: MatDisAssemble_MPIAIJ(mat);
5877: col = in[j];
5878: /* Reinitialize the variables required by MatSetValues_SeqAIJ_B_Private() */
5879: B = aij->B;
5880: b = (Mat_SeqAIJ*)B->data;
5881: bimax = b->imax; bi = b->i; bilen = b->ilen; bj = b->j;
5882: rp2 = bj + bi[row];
5883: ap2 = ba + bi[row];
5884: rmax2 = bimax[row];
5885: nrow2 = bilen[row];
5886: low2 = 0;
5887: high2 = nrow2;
5888: bm = aij->B->rmap->n;
5889: ba = b->a;
5890: }
5891: } else col = in[j];
5892: MatSetValues_SeqAIJ_B_Private(row,col,value,addv,im[i],in[j]);
5893: }
5894: }
5895: } else if (!aij->donotstash) {
5896: if (roworiented) {
5897: MatStashValuesRow_Private(&mat->stash,im[i],n,in,v+i*n,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5898: } else {
5899: MatStashValuesCol_Private(&mat->stash,im[i],n,in,v+i,m,(PetscBool)(ignorezeroentries && (addv == ADD_VALUES)));
5900: }
5901: }
5902: }
5903: }
5904: PetscFunctionReturnVoid();
5905: }